Загрузка видео...

Не удалось загрузить видео

На главную

Very quick comparison between Ornith 1.5 35B MoE and Qwen 3.8 27B Dense. 📣 Clearly it's not a fair one, but let's in any case see how it went. 397B download in progress! In the Videos below: - Brick SA -> Qwen 3.8 - Lego Streets -> Ornith 1.5...

ivanfioravanti's profile picture

Ivan Fioravanti ᯅ

42,009 subscribers

12,150 просмотров • 1 месяц назад •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

Alright guys, wow. Qwen 3.8 27B is basically Qwen 4. What we have here is one of the best one-shot shark survival games I've made with any model. I ran this same prompt through fable, and the biggest thing all models struggle with in this test is the top-down view of the sharks. Fable even had all of the fins reversed and details jumbled on the first turn. Before now, it was really a 2 to 8 turn test to get something final regardless of model. Qwen 3.8 27B just nailed this in one output, no harness at all, just purely the model thoroughly thinking before outputting the final result. On every asset used, different shark variants, everything, it all looks incredibly tight. The boat movement, shark speed and engagement, everything is so much better than I expected, and I was expecting a lot. Imagine what 8 turns would do in a harness? This confirms to me that while Qwen 3.8 27B thinks a lot, for the first time in a local model on the base weights, IT MIGHT ACTUALLY BE WORTH IT This is a Frontier Lab model result, and it ran on my 5090 at 60tps in Q5_K_M. I am just in complete astoundment at what Qwen has accomplished here, and I am incredibly grateful that this model is open source. The speed will also increase drastically with optimisations. This is just an incredible day. What a blessing. More to come!
4:28

Alright guys, wow. Qwen 3.8 27B is basically Qwen 4. What we have here is one of the best one-shot shark survival games I've made with any model. I ran this same prompt through fable, and the biggest thing all models struggle with in this test is the top-down view of the sharks. Fable even had all of the fins reversed and details jumbled on the first turn. Before now, it was really a 2 to 8 turn test to get something final regardless of model. Qwen 3.8 27B just nailed this in one output, no harness at all, just purely the model thoroughly thinking before outputting the final result. On every asset used, different shark variants, everything, it all looks incredibly tight. The boat movement, shark speed and engagement, everything is so much better than I expected, and I was expecting a lot. Imagine what 8 turns would do in a harness? This confirms to me that while Qwen 3.8 27B thinks a lot, for the first time in a local model on the base weights, IT MIGHT ACTUALLY BE WORTH IT This is a Frontier Lab model result, and it ran on my 5090 at 60tps in Q5_K_M. I am just in complete astoundment at what Qwen has accomplished here, and I am incredibly grateful that this model is open source. The speed will also increase drastically with optimisations. This is just an incredible day. What a blessing. More to come!

Kyle Hessling

24,917 просмотров • 1 месяц назад

MTPLX V2.10 is out! I put a lot of work into this one. Highlights: - Qwen 3.8 Next Support Added: Bare Speed (4 bit) and Optimised Speed (4 bit/8 bit) models added running at 50 - 80 TPS sustained at high contexts. (expect speed improvements in the coming days) - Qwen 3.8 Next mmap: The 32 GB n-gram table streams from SSD via mmap instead of RAM. Both models fit a 96 GB Mac with 0 speed impact. - Qwen 3.8 27B Decode & Prefill boost EVERYWHERE: +15% TPS at 3k context, +29% at 88k, +54% at 147k. Prefill +41%. - 48 GB Macs went from 3 to 4 TPS in swap @ 30k context to 33 TPS. - KV Cache has been solved!: 8 bit and 4 bit cache now barely impacts speed instead of a 50% decline. - Coding agents wall time optimisations: file task dropped from 150s to 44s, and mid-session first token went from ~2s to 0.11s. - CLI UX improvements + 34 bug fixes. My thoughts on Qwen Next? Incredible model. Enormously better at design, and creative tasks with scary good vision. It created what I think is the best flappy bird. maybe even better than the Fable version. It had various difficulties and skins, moving pillars, and upgrades. It also had the most memeified look. Incredibly impressed. However, it is slightly worse than 27B at having code just work straight away. But it's personality is also far better than 27B. Here is the Flappy bird it made below! Made with Qwen Next Optimized Speed one shot@ 57 TPS:
1:16

MTPLX V2.10 is out! I put a lot of work into this one. Highlights: - Qwen 3.8 Next Support Added: Bare Speed (4 bit) and Optimised Speed (4 bit/8 bit) models added running at 50 - 80 TPS sustained at high contexts. (expect speed improvements in the coming days) - Qwen 3.8 Next mmap: The 32 GB n-gram table streams from SSD via mmap instead of RAM. Both models fit a 96 GB Mac with 0 speed impact. - Qwen 3.8 27B Decode & Prefill boost EVERYWHERE: +15% TPS at 3k context, +29% at 88k, +54% at 147k. Prefill +41%. - 48 GB Macs went from 3 to 4 TPS in swap @ 30k context to 33 TPS. - KV Cache has been solved!: 8 bit and 4 bit cache now barely impacts speed instead of a 50% decline. - Coding agents wall time optimisations: file task dropped from 150s to 44s, and mid-session first token went from ~2s to 0.11s. - CLI UX improvements + 34 bug fixes. My thoughts on Qwen Next? Incredible model. Enormously better at design, and creative tasks with scary good vision. It created what I think is the best flappy bird. maybe even better than the Fable version. It had various difficulties and skins, moving pillars, and upgrades. It also had the most memeified look. Incredibly impressed. However, it is slightly worse than 27B at having code just work straight away. But it's personality is also far better than 27B. Here is the Flappy bird it made below! Made with Qwen Next Optimized Speed one shot@ 57 TPS:

Youssof Al Toukhi

21,126 просмотров • 25 дней назад

I'm really happy with Qwen3.8-27B ! All frustrations of not able to use Ox-Alpha and running out of credits for Glm-5.3, Grok turned me into Qwen-3.8-27b. Running locally on my 3090. This is the first fps game i'm creating with Qwen3.8-27b. I was after NMS vibe, alien planet, vast, flying drones, background music + sound effects. Without any external assets at all. With proper game like graphics settings panel 😍 And i'm really impressed, because all this is generated with my mixed quant for Qwen3.8-27B, which gives me more context + speed for a really good quality work. DSK harness at beginning and Pi agent towards the end cause i wanted to get this done fast. Full quant name: HF aj9o9/Qwen3.8-27B-GGUF file name: Qwen3.8-27B-gdn8-q6attn-iq3ffn.gguf Just like all of you, i also hate having to compact context, so higher context without ever needing to compromise quality is what i'm always after and this quant is created for that. Gives me 210K context at q8_0 kv + dflash2 Q4_K_M for speculative decoding. This result in 23GB vram usage.
3:11

I'm really happy with Qwen3.8-27B ! All frustrations of not able to use Ox-Alpha and running out of credits for Glm-5.3, Grok turned me into Qwen-3.8-27b. Running locally on my 3090. This is the first fps game i'm creating with Qwen3.8-27b. I was after NMS vibe, alien planet, vast, flying drones, background music + sound effects. Without any external assets at all. With proper game like graphics settings panel 😍 And i'm really impressed, because all this is generated with my mixed quant for Qwen3.8-27B, which gives me more context + speed for a really good quality work. DSK harness at beginning and Pi agent towards the end cause i wanted to get this done fast. Full quant name: HF aj9o9/Qwen3.8-27B-GGUF file name: Qwen3.8-27B-gdn8-q6attn-iq3ffn.gguf Just like all of you, i also hate having to compact context, so higher context without ever needing to compromise quality is what i'm always after and this quant is created for that. Gives me 210K context at q8_0 kv + dflash2 Q4_K_M for speculative decoding. This result in 23GB vram usage.

AJ

12,696 просмотров • 1 месяц назад

anything under deepseek v4 flash basically doesn't exist to me. with <200GB of vram, i'm getting ~800K context @ 200 tokens per second. It's not quite Cerebras fast, but damn. It is a joy to use, and its basically somewhere in the opus 4.7/4.8 range. It's incredibly good at browser control too. this model is a beast. w/ qwen 3.8 27B being multimodal though, it might just be worth switching to, especially if you want to run a lot of parallel agents will 300toks/s be possible? Pretty addicting let me tell you. Imagine a computer use model this fast. watch below, this is basically human speed.
0:28

anything under deepseek v4 flash basically doesn't exist to me. with <200GB of vram, i'm getting ~800K context @ 200 tokens per second. It's not quite Cerebras fast, but damn. It is a joy to use, and its basically somewhere in the opus 4.7/4.8 range. It's incredibly good at browser control too. this model is a beast. w/ qwen 3.8 27B being multimodal though, it might just be worth switching to, especially if you want to run a lot of parallel agents will 300toks/s be possible? Pretty addicting let me tell you. Imagine a computer use model this fast. watch below, this is basically human speed.

am.will

72,729 просмотров • 1 месяц назад

Qwen 3.8 Max beat Fable 5 at building 3D physics scenes for 7x cheaper! We gave two models the same task. Build three self-contained 3D scenes, each one HTML file with real physics that runs itself. Prompts: - A marble machine that lifts marbles up a wheel and drops them on a loop - A car factory assembly line - A sawmill cutting logs into planks Outputs: Qwen 3.8 Max: 46.6K tokens, $0.28 Fable 5: 38.7K tokens, $1.93 Qwen worked much harder on the details. Its textures and shadows look real. Fable 5 built a fast prototype that works and left it there. In the marble machine, Fable's marbles show up from nowhere at the top of the wheel. In Qwen's, you can see them go up. On the car line, Qwen's cars look much better and more real. Fable's sawmill is very strange. The saw looks wrong, and the log goes right through a beam before the saw cuts it. Atomic Chat will have day zero support to run Qwen 3.8 Max locally!
0:30

Sensitive content

Qwen 3.8 Max beat Fable 5 at building 3D physics scenes for 7x cheaper! We gave two models the same task. Build three self-contained 3D scenes, each one HTML file with real physics that runs itself. Prompts: - A marble machine that lifts marbles up a wheel and drops them on a loop - A car factory assembly line - A sawmill cutting logs into planks Outputs: Qwen 3.8 Max: 46.6K tokens, $0.28 Fable 5: 38.7K tokens, $1.93 Qwen worked much harder on the details. Its textures and shadows look real. Fable 5 built a fast prototype that works and left it there. In the marble machine, Fable's marbles show up from nowhere at the top of the wheel. In Qwen's, you can see them go up. On the car line, Qwen's cars look much better and more real. Fable's sawmill is very strange. The saw looks wrong, and the log goes right through a beam before the saw cuts it. Atomic Chat will have day zero support to run Qwen 3.8 Max locally!

atomic.chat

210,569 просмотров • 1 месяц назад

The VRAM barrier is officially dead. I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090. 21 tokens/sec decode. 364 t/s prefill. no mtp. no dflash. no kv cache quantization! We are running datacenter models on consumer hardware. Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs. ### The Benchmarks & Scaling # 1. Hybrid Offload (-ncmoe 40 @ 80k Context) Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling. ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40 Prefill: 383.85 t/s | Decode: 22.52 t/s Footprint: 23.85 GB VRAM | 97 GB RAM # 2. Full CPU MoE Offload (-cmoe @ 80k Context) Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090. llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe) Prefill: 355.72 t/s | Decode: 20.84 t/s Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM # 3. The 180,000 Context Run Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB # 4. The 250,000 Context Absolute Ceiling ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe Prefill: 364.29 t/s | Decode: 20.97 t/s Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM ### Key Insights: -b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s. -cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling. Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by Daniel Han: git clone && cd llama.cpp git fetch origin pull/27742/head:qwen-next && git checkout qwen-next cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF cmake --build build --config Release -j $(nproc) --target llama-server A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference. While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware. Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below. GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?
0:20

The VRAM barrier is officially dead. I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090. 21 tokens/sec decode. 364 t/s prefill. no mtp. no dflash. no kv cache quantization! We are running datacenter models on consumer hardware. Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs. ### The Benchmarks & Scaling # 1. Hybrid Offload (-ncmoe 40 @ 80k Context) Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling. ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40 Prefill: 383.85 t/s | Decode: 22.52 t/s Footprint: 23.85 GB VRAM | 97 GB RAM # 2. Full CPU MoE Offload (-cmoe @ 80k Context) Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090. llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe) Prefill: 355.72 t/s | Decode: 20.84 t/s Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM # 3. The 180,000 Context Run Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB # 4. The 250,000 Context Absolute Ceiling ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe Prefill: 364.29 t/s | Decode: 20.97 t/s Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM ### Key Insights: -b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s. -cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling. Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by Daniel Han: git clone && cd llama.cpp git fetch origin pull/27742/head:qwen-next && git checkout qwen-next cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF cmake --build build --config Release -j $(nproc) --target llama-server A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference. While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware. Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below. GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?

Alok

1,017,005 просмотров • 28 дней назад

Qwen 3.8 27B on hit 3.3x faster decode in 7 days. Here's what happened and what we're thinking next. Result (so far) Median decode speed increased from 26 tok/s to 87.9 tok/s on the verifier M5 Max (33 to 93.1 tok/s across the eight prompts), with prefill around 971.8 tok/s. This came out of a collective effort: 31 solvers across 67 improvements. Most of the recent ones run custom MTP heads that draft and accept ~3.9 tokens per round while still matching serial output exactly. Why this matters Beyond the performance itself, two things stand out to me. (1) Dense models on Apple Silicon were supposed to be the hard case. "Everyone knows Macs are slow at dense models." But watching the community take it from the usual baseline to >3x in seven days shows the low-hanging fruit was still there. (2) Open-weight models have been small and effective for a while. This is the first time one is small and frontier. Qwen 3.8 27B is an extremely strong dense model, comparable in capability to Opus 4.6 (Max). Running it at usable speed (>45 tok/s) is a step change for local AI users. What we improved about the challenge itself This is our second challenge, and we took the feedback from the Laguna track and rebuilt a few core pieces. - Speculative decoding (native MTP) was available and editable on day one instead of bolted on later. - Scoring became the median of eight independent prompt speedups over pure serial decode (anchored at 1.0, floor 0.90, ceiling 3.0), so no single fixture could dominate. - The leaderboard now ranks total contribution rather than just the current record holder. - Every submission gets automated screening for gaming before it scores. I really appreciate folks who's provided feedback. Naming a few that came to mind Ivan Fioravanti TheDavidTai Morgan McGuire poly Takeshi7 Steven Gumbii.Digital Tanishq Dubey Arjun Ram Andrey 🦃 Petrov tiny edge David Zhang Jaime Rader Surf and many others on slack! We also widened the editable surface to include the MTP head weights themselves, the full draft/verify loop, and a large set of the underlying Metal kernels. How we got to the 3x speedup Here's a summary from Grok. Much of it is beyond my understanding, but I expect people (and agents) smarter than I am can take these insights and apply them in other contexts. Custom MTP heads + adaptive draft policy People stopped treating the head as fixed and started training or editing it for higher acceptance under the exact verify constraints. Combined with per-round draft counts that can adapt (0 to 8), this is what pushed average accepted tokens from ~1-2 up to 3.9 on the top runs. Tighter verify-block and KV rollback paths The Swift session code for assembling the verify pass, snapshotting KV, and rolling back on rejects got cleaned up a lot. Small latency wins here compound once you're drafting ~four tokens at a time. Metal kernel work on the hot paths SDPA, the MoE gather GEMM, RoPE, RMSNorm, and a few of the smaller element-wise ops saw targeted edits. Most of the gains only show up once the verify width is high and the memory traffic pattern changes. Fidelity-preserving residual handling Several submissions improved how residuals and acceptance decisions are managed, so that higher draft depth doesn't quietly degrade the token match rate. The gates stayed strict: every emitted token still has to equal serial, so these were real engineering wins rather than score hacks. What's next for Qwen 3.8 27B MLX. We plan to keep the track live a bit longer, then switch to Qwen 3.8's MoE version (rumored to be 35B-A3B). Given the recent DFlash 2 announcement, we're also looking at whether we can support broader speculative methods. The current surface already supports a lot of experimentation. The main gaps are better upstreaming for local usage and clearer docs on how the benchmark and verifier work. Multiplatform. In parallel, we're experimenting with running a similar effort around CUDA for Qwen 3.8 27B. A lot of people have asked for this, since the two communities overlap quite a bit. Our goal is to ship the CUDA version next week. We'd also love to partner with Qwen on it. If anyone has a connection there, please introduce us, and we'll see if they're down to match a bounty with us to push this out. What's most useful for the broader MLX community The improvements from the challenge are already upstreamed inside Darkbloom, and we're seeing ~2x faster decode in our production traffic for Qwen. Outside the challenge itself, something I've been thinking about deeply, and that a few community members have raised, is how to make these results useful to more people. There are many individual efforts happening across the MLX community, and honestly, the more I dig in, the more confused I get by the overlapping libraries and concepts. I'm sure I'm not alone, and newcomers probably feel the same. That's no one's fault, just the growing pains of an open source community. I don't expect I'm gonna come up with the answer, but I'd love to learn more about what different folks are working on and how they're thinking about their roadmaps. I'll share what I learn along the way, and hopefully someone smarter than me can turn it into a proposal for us to rally around.
0:27

Qwen 3.8 27B on hit 3.3x faster decode in 7 days. Here's what happened and what we're thinking next. Result (so far) Median decode speed increased from 26 tok/s to 87.9 tok/s on the verifier M5 Max (33 to 93.1 tok/s across the eight prompts), with prefill around 971.8 tok/s. This came out of a collective effort: 31 solvers across 67 improvements. Most of the recent ones run custom MTP heads that draft and accept ~3.9 tokens per round while still matching serial output exactly. Why this matters Beyond the performance itself, two things stand out to me. (1) Dense models on Apple Silicon were supposed to be the hard case. "Everyone knows Macs are slow at dense models." But watching the community take it from the usual baseline to >3x in seven days shows the low-hanging fruit was still there. (2) Open-weight models have been small and effective for a while. This is the first time one is small and frontier. Qwen 3.8 27B is an extremely strong dense model, comparable in capability to Opus 4.6 (Max). Running it at usable speed (>45 tok/s) is a step change for local AI users. What we improved about the challenge itself This is our second challenge, and we took the feedback from the Laguna track and rebuilt a few core pieces. - Speculative decoding (native MTP) was available and editable on day one instead of bolted on later. - Scoring became the median of eight independent prompt speedups over pure serial decode (anchored at 1.0, floor 0.90, ceiling 3.0), so no single fixture could dominate. - The leaderboard now ranks total contribution rather than just the current record holder. - Every submission gets automated screening for gaming before it scores. I really appreciate folks who's provided feedback. Naming a few that came to mind Ivan Fioravanti TheDavidTai Morgan McGuire poly Takeshi7 Steven Gumbii.Digital Tanishq Dubey Arjun Ram Andrey 🦃 Petrov tiny edge David Zhang Jaime Rader Surf and many others on slack! We also widened the editable surface to include the MTP head weights themselves, the full draft/verify loop, and a large set of the underlying Metal kernels. How we got to the 3x speedup Here's a summary from Grok. Much of it is beyond my understanding, but I expect people (and agents) smarter than I am can take these insights and apply them in other contexts. Custom MTP heads + adaptive draft policy People stopped treating the head as fixed and started training or editing it for higher acceptance under the exact verify constraints. Combined with per-round draft counts that can adapt (0 to 8), this is what pushed average accepted tokens from ~1-2 up to 3.9 on the top runs. Tighter verify-block and KV rollback paths The Swift session code for assembling the verify pass, snapshotting KV, and rolling back on rejects got cleaned up a lot. Small latency wins here compound once you're drafting ~four tokens at a time. Metal kernel work on the hot paths SDPA, the MoE gather GEMM, RoPE, RMSNorm, and a few of the smaller element-wise ops saw targeted edits. Most of the gains only show up once the verify width is high and the memory traffic pattern changes. Fidelity-preserving residual handling Several submissions improved how residuals and acceptance decisions are managed, so that higher draft depth doesn't quietly degrade the token match rate. The gates stayed strict: every emitted token still has to equal serial, so these were real engineering wins rather than score hacks. What's next for Qwen 3.8 27B MLX. We plan to keep the track live a bit longer, then switch to Qwen 3.8's MoE version (rumored to be 35B-A3B). Given the recent DFlash 2 announcement, we're also looking at whether we can support broader speculative methods. The current surface already supports a lot of experimentation. The main gaps are better upstreaming for local usage and clearer docs on how the benchmark and verifier work. Multiplatform. In parallel, we're experimenting with running a similar effort around CUDA for Qwen 3.8 27B. A lot of people have asked for this, since the two communities overlap quite a bit. Our goal is to ship the CUDA version next week. We'd also love to partner with Qwen on it. If anyone has a connection there, please introduce us, and we'll see if they're down to match a bounty with us to push this out. What's most useful for the broader MLX community The improvements from the challenge are already upstreamed inside Darkbloom, and we're seeing ~2x faster decode in our production traffic for Qwen. Outside the challenge itself, something I've been thinking about deeply, and that a few community members have raised, is how to make these results useful to more people. There are many individual efforts happening across the MLX community, and honestly, the more I dig in, the more confused I get by the overlapping libraries and concepts. I'm sure I'm not alone, and newcomers probably feel the same. That's no one's fault, just the growing pains of an open source community. I don't expect I'm gonna come up with the answer, but I'd love to learn more about what different folks are working on and how they're thinking about their roadmaps. I'll share what I learn along the way, and hopefully someone smarter than me can turn it into a proposal for us to rally around.

Kydo

32,945 просмотров • 1 месяц назад

First impressions on Muse Glimmer! It's incredibly fast for a dense model, currently running an average of 208tps with a max of 274tps on a single 5090 with their DFLASH config. Comparatively, though, both using Open Code, Qwopus Coder (with thinking off) produced a much better shark survival game than the one I got from Glimmer. Meta's new dense model is currently just lacking some HTML canvas taste, but this is something that can be added via SFT as long as the model is stable and capable from a back-end programming perspective. And it seems to be, without a doubt. The big kicker here is that I ran this at extra high thinking, and it did not take long at all to run. Our current local leader, Qwen 27B 3.6, has a tendency to overthink, but with glimmer, that is not the case. Right now, my recommendation for general local programming (Apps, Games, Websites, Visual Tools) in this class is still Qwopus Coder with thinking disabled, or Qwopus Fusion with thinking enabled. Of course Shark Survival is a very basic domain-specific test, but I find that the result scales very well across many domains. If we're going to be shipping apps generated entirely locally, visual taste is somewhat of a bare minimum requirement, solely in my opinion, and Qwen's models in this class offer significantly more at the moment. That's actually why I initially started getting into finetuning with Qwen 3.5, they were the first base that was able to do really good front-end with some opus-trace fine-tuning. Qwen 3.6 has taste even in the base model, and we know Qwen 3.8 is going to blow us all away! Regardless, this looks like a very tempting new base model. As a first offering from Meta in this class for a long time, I am incredibly impressed and elated to have it. We now finally have a proper Single GPU frontier race, instead of us just begging Qwen for more releases. Single GPU open frontier model race is a VERY good thing. Please keep pushing Meta
1:20

First impressions on Muse Glimmer! It's incredibly fast for a dense model, currently running an average of 208tps with a max of 274tps on a single 5090 with their DFLASH config. Comparatively, though, both using Open Code, Qwopus Coder (with thinking off) produced a much better shark survival game than the one I got from Glimmer. Meta's new dense model is currently just lacking some HTML canvas taste, but this is something that can be added via SFT as long as the model is stable and capable from a back-end programming perspective. And it seems to be, without a doubt. The big kicker here is that I ran this at extra high thinking, and it did not take long at all to run. Our current local leader, Qwen 27B 3.6, has a tendency to overthink, but with glimmer, that is not the case. Right now, my recommendation for general local programming (Apps, Games, Websites, Visual Tools) in this class is still Qwopus Coder with thinking disabled, or Qwopus Fusion with thinking enabled. Of course Shark Survival is a very basic domain-specific test, but I find that the result scales very well across many domains. If we're going to be shipping apps generated entirely locally, visual taste is somewhat of a bare minimum requirement, solely in my opinion, and Qwen's models in this class offer significantly more at the moment. That's actually why I initially started getting into finetuning with Qwen 3.5, they were the first base that was able to do really good front-end with some opus-trace fine-tuning. Qwen 3.6 has taste even in the base model, and we know Qwen 3.8 is going to blow us all away! Regardless, this looks like a very tempting new base model. As a first offering from Meta in this class for a long time, I am incredibly impressed and elated to have it. We now finally have a proper Single GPU frontier race, instead of us just begging Qwen for more releases. Single GPU open frontier model race is a VERY good thing. Please keep pushing Meta

Kyle Hessling

17,485 просмотров • 1 месяц назад

i pointed hermes agent at nvidia's nemotron cascade 2 30B-A3B on a single RTX 3090 24GB. IQ4_XS quant by bartowski, 187 tok/s, 625K context. had it discover its own hardware, create an identity file, then build a full GPU marketplace UI from a single prompt. it one shotted it. first attempt no iteration. qwen 3.5 35B-A3B on the same hardware same 3090 24GB took an iteration to recover from a blank screen on the same type of build. 24 days between these two models releasing. same active parameters, completely different architectures and cascade 2 through hermes agent just keeps going. this model goes on and on. feast your eyes. more iterations and tests dropping soon. nvidia really cooked. no special flags needed. nvidia optimized this mamba MoE so well it just runs. flash attention auto enabled, context auto allocated. the model does the work not the config. but i compiled llama.cpp from source and i'm not sure how it performs on other engines. if you ran nemotron on any hardware drop your numbers below. RTX, AMD, Mac, whatever. model, quant, tok/s, engine. i want to see if it holds everywhere or just on llama.cpp.
1:31

i pointed hermes agent at nvidia's nemotron cascade 2 30B-A3B on a single RTX 3090 24GB. IQ4_XS quant by bartowski, 187 tok/s, 625K context. had it discover its own hardware, create an identity file, then build a full GPU marketplace UI from a single prompt. it one shotted it. first attempt no iteration. qwen 3.5 35B-A3B on the same hardware same 3090 24GB took an iteration to recover from a blank screen on the same type of build. 24 days between these two models releasing. same active parameters, completely different architectures and cascade 2 through hermes agent just keeps going. this model goes on and on. feast your eyes. more iterations and tests dropping soon. nvidia really cooked. no special flags needed. nvidia optimized this mamba MoE so well it just runs. flash attention auto enabled, context auto allocated. the model does the work not the config. but i compiled llama.cpp from source and i'm not sure how it performs on other engines. if you ran nemotron on any hardware drop your numbers below. RTX, AMD, Mac, whatever. model, quant, tok/s, engine. i want to see if it holds everywhere or just on llama.cpp.

Sudo su

71,470 просмотров • 6 месяцев назад

Darkbloom is the first model provider to support Ternary Bonsai 2 27B -- concentrated intelligence that fits on your phone. Try it now: First 250 users get 100 million free tokens; PrismML's new flagship model: - a ternary compression of Qwen3.8 27B at 2bits per parameter. - 8.5 GB total, 5x smaller than original - keeps 98.2% of the Qwen's FP16 benchmark performance. - 75% cheaper than Qwen 27B. Qwen3.8 27B already operates comparably with Opus 4.6 and 5.6 Luna on certain tasks. This one does it in the memory of a phone. 262K context, image input, Apache 2.0. From our first run on the network, on a single M5 Max with no caching: - 35 tok/s decode at 1K context, - 31 tok/s at 10K, - 19 tok/s at 50K. But we expect more performance gain coming in a few weeks! That's a full 27B reasoning model running comfortably on any Mac. 1,000+ Macs are serving on Darkbloom right now. Go try it out!! Thank you to Babak Hassibi Sahin Lale Omead Pooladzandi Reza Sadri Tushar Bansal Evan Walters and the whole PrismML team. This is exactly the kind of model Darkbloom was built for. You can read the essay by Bonsai on Why Local AI Matters:
1:20

Darkbloom is the first model provider to support Ternary Bonsai 2 27B -- concentrated intelligence that fits on your phone. Try it now: First 250 users get 100 million free tokens; PrismML's new flagship model: - a ternary compression of Qwen3.8 27B at 2bits per parameter. - 8.5 GB total, 5x smaller than original - keeps 98.2% of the Qwen's FP16 benchmark performance. - 75% cheaper than Qwen 27B. Qwen3.8 27B already operates comparably with Opus 4.6 and 5.6 Luna on certain tasks. This one does it in the memory of a phone. 262K context, image input, Apache 2.0. From our first run on the network, on a single M5 Max with no caching: - 35 tok/s decode at 1K context, - 31 tok/s at 10K, - 19 tok/s at 50K. But we expect more performance gain coming in a few weeks! That's a full 27B reasoning model running comfortably on any Mac. 1,000+ Macs are serving on Darkbloom right now. Go try it out!! Thank you to Babak Hassibi Sahin Lale Omead Pooladzandi Reza Sadri Tushar Bansal Evan Walters and the whole PrismML team. This is exactly the kind of model Darkbloom was built for. You can read the essay by Bonsai on Why Local AI Matters:

Gajesh

46,273 просмотров • 5 дней назад

hey if you're thinking about running qwopus (the claude opus distilled qwen 3.5 27B) as a coding agent, this might save you a few hours. i tested both the base and the distilled version on the same hardware. single RTX 3090. same prompt. same context. same everything. the only variable was the model weights. base qwen 3.5 27B built octopus invaders in 13 minutes. 1,827 lines across 11 files. zero steering. one scope bug that took 2 lines to fix. game ran. qwopus couldn't finish the same task. enemies overlapping on screen. bullets not firing. controls worked but the game was broken. i had to steer it multiple times and it still didn't produce a playable result. both run at 35 tok/s. both use thinking mode. the distilled version actually has better jinja compatibility and doesn't stall midtask like base does on claude code. for conversation and reasoning it feels sharper. but for multifile autonomous coding where the model needs to coordinate 10+ files without losing track, base wins and it's not close. distillation compresses reasoning patterns but seems to lose precision on complex coordination. the model "thinks" well but can't hold the full picture across files the way base can. tested on opencode (base) and claude code (both). next up is hermes agent framework on base. same hardware. same prompt. comparing agents now, not just models. video below. first half is the distilled model's broken game. second half is what base built on the same 3090. judge for yourself.
0:20

hey if you're thinking about running qwopus (the claude opus distilled qwen 3.5 27B) as a coding agent, this might save you a few hours. i tested both the base and the distilled version on the same hardware. single RTX 3090. same prompt. same context. same everything. the only variable was the model weights. base qwen 3.5 27B built octopus invaders in 13 minutes. 1,827 lines across 11 files. zero steering. one scope bug that took 2 lines to fix. game ran. qwopus couldn't finish the same task. enemies overlapping on screen. bullets not firing. controls worked but the game was broken. i had to steer it multiple times and it still didn't produce a playable result. both run at 35 tok/s. both use thinking mode. the distilled version actually has better jinja compatibility and doesn't stall midtask like base does on claude code. for conversation and reasoning it feels sharper. but for multifile autonomous coding where the model needs to coordinate 10+ files without losing track, base wins and it's not close. distillation compresses reasoning patterns but seems to lose precision on complex coordination. the model "thinks" well but can't hold the full picture across files the way base can. tested on opencode (base) and claude code (both). next up is hermes agent framework on base. same hardware. same prompt. comparing agents now, not just models. video below. first half is the distilled model's broken game. second half is what base built on the same 3090. judge for yourself.

Sudo su

45,052 просмотров • 6 месяцев назад

Qwen3.8-27B running at full BF16 on a free Kaggle TPU is kind of ridiculous. No quantization. No tiny context window. No expensive GPU instance. Just Qwen3.8-27B running on a Kaggle TPU v5e-8. The reported numbers: ~130 tok/s decode ~10,000 tok/s prefill 262K context That prefill number is especially wild. You can throw a huge amount of code or context at the model and ingest it extremely quickly, while still getting around 130 tokens per second during generation. And because it’s running in full BF16, you’re not relying on an aggressive quant just to make the model fit. But the really interesting part isn’t even the raw throughput. You can expose it as an OpenAI-compatible endpoint. That means you can plug the model into tools that already understand OpenAI-style APIs. Claude Code. Codex. OpenCode. And other compatible clients. So the workflow becomes pretty simple: Spin up the Qwen3.8-27B endpoint on Kaggle. Point your coding tool at the API. And suddenly you have a 27B coding model sitting behind the same interface you’d normally use for hosted models. The 262K context is also a huge deal for agentic coding. Large repositories can fit into a single context. Long conversations don’t need to be constantly trimmed. And tools can feed much more information back to the model without hitting a tiny context ceiling. The fact that this can be built around a free TPU environment is what makes this especially interesting. We’re getting to a point where experimenting with serious open models doesn’t always require owning a $2,000 GPU or paying for a large cloud instance. Free compute + open weights + an OpenAI-compatible API + existing coding agents. That’s a pretty powerful combination. Qwen3.8-27B is already an interesting model. Running the full BF16 version at ~130 tok/s with 262K context on free Kaggle TPU compute makes it a lot more interesting.
0:48

Qwen3.8-27B running at full BF16 on a free Kaggle TPU is kind of ridiculous. No quantization. No tiny context window. No expensive GPU instance. Just Qwen3.8-27B running on a Kaggle TPU v5e-8. The reported numbers: ~130 tok/s decode ~10,000 tok/s prefill 262K context That prefill number is especially wild. You can throw a huge amount of code or context at the model and ingest it extremely quickly, while still getting around 130 tokens per second during generation. And because it’s running in full BF16, you’re not relying on an aggressive quant just to make the model fit. But the really interesting part isn’t even the raw throughput. You can expose it as an OpenAI-compatible endpoint. That means you can plug the model into tools that already understand OpenAI-style APIs. Claude Code. Codex. OpenCode. And other compatible clients. So the workflow becomes pretty simple: Spin up the Qwen3.8-27B endpoint on Kaggle. Point your coding tool at the API. And suddenly you have a 27B coding model sitting behind the same interface you’d normally use for hosted models. The 262K context is also a huge deal for agentic coding. Large repositories can fit into a single context. Long conversations don’t need to be constantly trimmed. And tools can feed much more information back to the model without hitting a tiny context ceiling. The fact that this can be built around a free TPU environment is what makes this especially interesting. We’re getting to a point where experimenting with serious open models doesn’t always require owning a $2,000 GPU or paying for a large cloud instance. Free compute + open weights + an OpenAI-compatible API + existing coding agents. That’s a pretty powerful combination. Qwen3.8-27B is already an interesting model. Running the full BF16 version at ~130 tok/s with 262K context on free Kaggle TPU compute makes it a lot more interesting.

FHILY👑

35,794 просмотров • 17 дней назад

first impressions of qwen 3.5 27B dense on a single RTX 3090. 35 tok/s. from 4K all the way to 300K+ context. no speed drop. hermes 4.3 started at 35 and degraded to 15 as context filled. qwen dense holds. MoE held 112 flat. 3x faster but only 3B of 35B active per token. architecture tradeoff. Q4_K_M on 16.7GB. native context 262K. pushed past training limit to 376K before VRAM ceiling on 24GB. tried q8 KV cache at 262K, speed collapsed to 11 tok/s. q4_0 KV is the sweet spot. flash attention mandatory. built in reasoning mode. the model thinks step by step before it answers. full chain of thought surviving Q4 quant. 1,799+ token thinking chains with self correction loops. on a single consumer GPU. gave it one prompt: "build a realtime particle galaxy simulation in one HTML file." 3,340 tokens. 95 seconds. one shot. ran on first load. full reasoning and coding in the video below. optimal config if you want to skip the hours of testing: llama-server -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 this is just the warmup. octopus invaders is next: 10 files, 3,400+ lines, zero steering. the prompt hermes quit at 22%. already more impressed than expected. full results coming soon.
2:02

first impressions of qwen 3.5 27B dense on a single RTX 3090. 35 tok/s. from 4K all the way to 300K+ context. no speed drop. hermes 4.3 started at 35 and degraded to 15 as context filled. qwen dense holds. MoE held 112 flat. 3x faster but only 3B of 35B active per token. architecture tradeoff. Q4_K_M on 16.7GB. native context 262K. pushed past training limit to 376K before VRAM ceiling on 24GB. tried q8 KV cache at 262K, speed collapsed to 11 tok/s. q4_0 KV is the sweet spot. flash attention mandatory. built in reasoning mode. the model thinks step by step before it answers. full chain of thought surviving Q4 quant. 1,799+ token thinking chains with self correction loops. on a single consumer GPU. gave it one prompt: "build a realtime particle galaxy simulation in one HTML file." 3,340 tokens. 95 seconds. one shot. ran on first load. full reasoning and coding in the video below. optimal config if you want to skip the hours of testing: llama-server -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 this is just the warmup. octopus invaders is next: 10 files, 3,400+ lines, zero steering. the prompt hermes quit at 22%. already more impressed than expected. full results coming soon.

Sudo su

122,051 просмотров • 6 месяцев назад

hey here is the final result of octopus invaders on nvidia's flagship at full precision. nemotron super 120B on 2x H200 NVL. BF16 unquantized. 287GB of VRAM. hermes agent as the harness. 60 tok/s. first try it autonomously coded for 6 minutes straight. created 11 files. correct project structure. correct load order. started the server. i opened the browser and the result was a blank screen. i did not give up. second try i gave it a precise list of bugs and things to fix. it went back in for another 3 minutes. patched the code. served it again. still blank. so i did what any sane person would do. third try i just said the screen is blank, test it and fix it yourself. and this is where nemotron showed what it actually is. it became a debugger. you can see it in the video. realtime CSS test squares, red screen flashes, hermes agent browser tools, inspecting its own output. it built the parallax background with planets and comets. it rendered a rocket ship that tracks your mouse with fire and bullet physics. the aesthetic is real. but no enemies spawn. no collision. not playable. what surprised me is qwen 27B one shotted this exact game on a single RTX 3090 at Q4 quant. and here is nvidia's flagship at full precision on enterprise hardware needing 3 tries and still not getting there. that makes my hope high for the undisputed qwen 122B which is about to face the same test next. same hardware. same prompt and same harness. lets see if it one shots or not. full session in the video. no cuts. 5x speed.
3:36

hey here is the final result of octopus invaders on nvidia's flagship at full precision. nemotron super 120B on 2x H200 NVL. BF16 unquantized. 287GB of VRAM. hermes agent as the harness. 60 tok/s. first try it autonomously coded for 6 minutes straight. created 11 files. correct project structure. correct load order. started the server. i opened the browser and the result was a blank screen. i did not give up. second try i gave it a precise list of bugs and things to fix. it went back in for another 3 minutes. patched the code. served it again. still blank. so i did what any sane person would do. third try i just said the screen is blank, test it and fix it yourself. and this is where nemotron showed what it actually is. it became a debugger. you can see it in the video. realtime CSS test squares, red screen flashes, hermes agent browser tools, inspecting its own output. it built the parallax background with planets and comets. it rendered a rocket ship that tracks your mouse with fire and bullet physics. the aesthetic is real. but no enemies spawn. no collision. not playable. what surprised me is qwen 27B one shotted this exact game on a single RTX 3090 at Q4 quant. and here is nvidia's flagship at full precision on enterprise hardware needing 3 tries and still not getting there. that makes my hope high for the undisputed qwen 122B which is about to face the same test next. same hardware. same prompt and same harness. lets see if it one shots or not. full session in the video. no cuts. 5x speed.

Sudo su

11,011 просмотров • 5 месяцев назад

🥊 Time for a new fight Qwen3.5 397B A17B vs Qwen3.6 27B 🌸 "Cherry Blossom" (↓ prompt below) Using OpenRouter for 397B Running 27B locally on a single RTX 5080 Wow! 🤯 I'm sure this one is going to divide opinions. 397B feels cinematic with the flare and shadows it added, while 27B is just beautiful; it has a completely different color palette and the leaves are falling from the branches as asked. It's really tough to decide which model is the best. I was about to call it a tie, but then I remembered that 397B needs 180GB VRAM to run locally 😅 Don't get me wrong the model is amazing, but right now the VRAM it requires doesn't feel justified. It's simply not 10× better than 27B. We'll have to dig deeper with more prompts to be sure, let's see what each one has in store for us. What do you think?
0:30

🥊 Time for a new fight Qwen3.5 397B A17B vs Qwen3.6 27B 🌸 "Cherry Blossom" (↓ prompt below) Using OpenRouter for 397B Running 27B locally on a single RTX 5080 Wow! 🤯 I'm sure this one is going to divide opinions. 397B feels cinematic with the flare and shadows it added, while 27B is just beautiful; it has a completely different color palette and the leaves are falling from the branches as asked. It's really tough to decide which model is the best. I was about to call it a tie, but then I remembered that 397B needs 180GB VRAM to run locally 😅 Don't get me wrong the model is amazing, but right now the VRAM it requires doesn't feel justified. It's simply not 10× better than 27B. We'll have to dig deeper with more prompts to be sure, let's see what each one has in store for us. What do you think?

left curve dev

21,224 просмотров • 4 месяцев назад

When using GPT-5.5, it is instantly noticeable how much more powerful it is. In Codex, I gave it a very complex prompt to create London Toy Railway with landmarks and seasons - it did an excellent job in one shot. In the second half of the video you see GPT-5.4 - it was also not bad, but very clearly worse. GPT-5.5's generation is far more ambitious, coherent and with fewer errors. This is obviously a toy example, but I've used it on much more complex real tasks, including a complex app migration and a new hard workflow - it has been working away for many hours without getting stumped. I'm getting more and more addicted to this stuff with every model release.
0:49

When using GPT-5.5, it is instantly noticeable how much more powerful it is. In Codex, I gave it a very complex prompt to create London Toy Railway with landmarks and seasons - it did an excellent job in one shot. In the second half of the video you see GPT-5.4 - it was also not bad, but very clearly worse. GPT-5.5's generation is far more ambitious, coherent and with fewer errors. This is obviously a toy example, but I've used it on much more complex real tasks, including a complex app migration and a new hard workflow - it has been working away for many hours without getting stumped. I'm getting more and more addicted to this stuff with every model release.

Peter Gostev

263,046 просмотров • 5 месяцев назад

Run Updated Gemma 4 26B A4B QAT (MoE) with Vision at 25 tokens/sec and massive 120k context window on a single RTX 4060 (8 GB VRAM + 16 GB RAM Only!!) Yesterday I pushed Gemma 4 26B A4B QAT to 250k context on a single RTX 4060 using nothing but Q8 KV cache and optimized -b and -ub flags for higher prefill throughput. Today I stacked Multi Token Prediction (MTP) self speculative decoding AND the vision projector (mmproj) on top of that same card, same batch size optimization, same $250 GPU and pushed it until it broke, then found the fix. All text only runs consist of a 28k prompt. vision runs consist of 28k text prompt + an image. # 1. MTP alone. near free decode speed, no catch MTP draft assistant is a separate small model (MTP heads are backed into the main model itself for the qwen 3.5+ models but its a separate small model for gemma 4 series), 240 MB gguf 80k ctx: Prefill 510 t/s | Decode 29.5 t/s 120k ctx: Prefill 433 t/s | Decode 29 t/s 180k ctx: Prefill 240 t/s | Decode 24.9 t/s 250k ctx: Prefill 63 t/s | Decode 13 t/s llama.cpp flags: m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf-c 180000 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 # 2. Add vision on top. the tax you actually pay the vision projector gguf is about 1.1 GBs 80k ctx: Prefill 360 t/s | Decode 25.4 t/s 120k ctx: Prefill 230 t/s | Decode 23.8 t/s 180k ctx (Q8 KV): Prefill 75 t/s | Decode 12.5 t/s - cliff flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf -c 80000 --port 8080 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 --mmproj mmproj-F16.gguf # 3. The fix if you want to run vision over 120k context: swap Q8 KV for Q4 KV past 120k Stack MTP + vision + Q8 KV past 120k context and you hit a wall. draft model overhead plus KV pressure tanks everything. Drop to Q4 KV and the wall disappears: 180k ctx (Q4 KV): Prefill 220 t/s | Decode 25.5 t/s -ctk q4_0 -ctv q4_0 --mmproj mmproj-F16.gguf (rest same as above) Bottom line: MTP gives you a near free +20-30% decode boost up to 120k context. Past that, it's fighting your VRAM, not helping and if vision is loaded too, Q4 KV isn't optional past 120k, it's mandatory. 30% boost is model and card specific, MTP boosted decode 2x for gemma 4 31b on a single rtx 4090. Same 8GB card. Same $250 GPU. Multimodal, speculative decoding, 180k usable context, zero upgrades. You gotta try this if you have a single NVIDIA RTX 3050, 3060, 3070, 4050, 4060, 5050 or 5060. You can try it with a 6 GB VRAM card as well but you will have to lower the context window. Hugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below. Which models are you running on your 6/8/12GB cards with MTP?
0:20

Run Updated Gemma 4 26B A4B QAT (MoE) with Vision at 25 tokens/sec and massive 120k context window on a single RTX 4060 (8 GB VRAM + 16 GB RAM Only!!) Yesterday I pushed Gemma 4 26B A4B QAT to 250k context on a single RTX 4060 using nothing but Q8 KV cache and optimized -b and -ub flags for higher prefill throughput. Today I stacked Multi Token Prediction (MTP) self speculative decoding AND the vision projector (mmproj) on top of that same card, same batch size optimization, same $250 GPU and pushed it until it broke, then found the fix. All text only runs consist of a 28k prompt. vision runs consist of 28k text prompt + an image. # 1. MTP alone. near free decode speed, no catch MTP draft assistant is a separate small model (MTP heads are backed into the main model itself for the qwen 3.5+ models but its a separate small model for gemma 4 series), 240 MB gguf 80k ctx: Prefill 510 t/s | Decode 29.5 t/s 120k ctx: Prefill 433 t/s | Decode 29 t/s 180k ctx: Prefill 240 t/s | Decode 24.9 t/s 250k ctx: Prefill 63 t/s | Decode 13 t/s llama.cpp flags: m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf-c 180000 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 # 2. Add vision on top. the tax you actually pay the vision projector gguf is about 1.1 GBs 80k ctx: Prefill 360 t/s | Decode 25.4 t/s 120k ctx: Prefill 230 t/s | Decode 23.8 t/s 180k ctx (Q8 KV): Prefill 75 t/s | Decode 12.5 t/s - cliff flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf -c 80000 --port 8080 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 --mmproj mmproj-F16.gguf # 3. The fix if you want to run vision over 120k context: swap Q8 KV for Q4 KV past 120k Stack MTP + vision + Q8 KV past 120k context and you hit a wall. draft model overhead plus KV pressure tanks everything. Drop to Q4 KV and the wall disappears: 180k ctx (Q4 KV): Prefill 220 t/s | Decode 25.5 t/s -ctk q4_0 -ctv q4_0 --mmproj mmproj-F16.gguf (rest same as above) Bottom line: MTP gives you a near free +20-30% decode boost up to 120k context. Past that, it's fighting your VRAM, not helping and if vision is loaded too, Q4 KV isn't optional past 120k, it's mandatory. 30% boost is model and card specific, MTP boosted decode 2x for gemma 4 31b on a single rtx 4090. Same 8GB card. Same $250 GPU. Multimodal, speculative decoding, 180k usable context, zero upgrades. You gotta try this if you have a single NVIDIA RTX 3050, 3060, 3070, 4050, 4060, 5050 or 5060. You can try it with a 6 GB VRAM card as well but you will have to lower the context window. Hugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below. Which models are you running on your 6/8/12GB cards with MTP?

Alok

16,405 просмотров • 2 месяцев назад

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.
0:33

Qwen3.8-Flash-Next is still going strong at 364.7K tokens of context on an M5 Max. And this isn’t just a static long-context test. The model was reasoning about how to speed up its own workflow while using tools, and the tool calls kept working without misses. Setup: • Qwen3.8-Flash-Next • M5 Max • 128GB unified memory • MLX-Serve PR #363 • OpenCode 2 • 364.7K context The interesting part isn’t simply getting hundreds of thousands of tokens into memory. It’s what happens once the context gets this large. Long-context inference usually comes with a painful tradeoff. As the KV cache grows, memory pressure increases and generation can slow down. But this setup is still pushing through 364K tokens while maintaining a usable agent workflow. The model can reason, call tools, inspect results, continue working, and keep the session moving. And the tool calls reportedly haven’t missed so far. That’s important for agentic coding. A huge context window is only useful if the model can actually operate reliably inside it. A 400K-token context that constantly breaks tool calls isn’t very useful. A 364K session that can keep reasoning and executing tools is a different story. And the test isn’t finished yet. The current run is approaching 400K tokens, with the expectation that it can keep going. This is also another interesting example of why Apple Silicon keeps showing up in local LLM experiments. The M5 Max’s unified memory gives a large model and its growing KV cache access to one shared memory pool. With MLX-Serve continuing to improve, these machines are becoming surprisingly capable long-context inference boxes. The bigger takeaway: Context length is becoming a workload, not just a model specification. Running a model at 256K is one thing. Keeping an agent alive at 300K+ while it reasons and uses tools is much more interesting. And Qwen3.8-Flash-Next is showing that this can be pushed surprisingly far on a single 128GB Mac. 364.7K and counting. Next stop: 400K.

FHILY👑

39,982 просмотров • 16 дней назад

Qwen 3.8 Flash Next (125B-A6B) now runs natively in mlx-serve on Apple Silicon. No Python, Zig + Metal. (Preview Release) M4 Max, 4-bit pack (~75 GB resident): - ~70 tok/s serial, ~98 tok/s with MTP when tested using `npx llmprobe` - prefill ~700 tok/s, sparse attention past 2k tokens with fused kernels - images and video input supported Hyper-connections, n-gram embeddings gathered straight from disk, QSA block selection, 5 new Metal kernels. All open source, parity-tested against the HF reference. Kernel work: - hyper-connection read: 3 Metal kernels, split-K, write folded into the next read - GDN decode: 2 fused dispatches per layer - hd-256 attention: mask arm, sparse prefill + verify stay fused - all greedy-equivalent to the unfused path If you port any of this to your engine, a mention would be nice. MLX-Serve Preview 26.8.11 build: 4bit Speed Pack: 4-8bit Quality Pack: As for the model, i'm not sure yet, it feels verbose, and Opus like. Thinks a lot. Goes on long self testing loops/talking with itself, but need more time with it. It seems like a good model for agentic things.
0:25

Qwen 3.8 Flash Next (125B-A6B) now runs natively in mlx-serve on Apple Silicon. No Python, Zig + Metal. (Preview Release) M4 Max, 4-bit pack (~75 GB resident): - ~70 tok/s serial, ~98 tok/s with MTP when tested using `npx llmprobe` - prefill ~700 tok/s, sparse attention past 2k tokens with fused kernels - images and video input supported Hyper-connections, n-gram embeddings gathered straight from disk, QSA block selection, 5 new Metal kernels. All open source, parity-tested against the HF reference. Kernel work: - hyper-connection read: 3 Metal kernels, split-K, write folded into the next read - GDN decode: 2 fused dispatches per layer - hd-256 attention: mask arm, sparse prefill + verify stay fused - all greedy-equivalent to the unfused path If you port any of this to your engine, a mention would be nice. MLX-Serve Preview 26.8.11 build: 4bit Speed Pack: 4-8bit Quality Pack: As for the model, i'm not sure yet, it feels verbose, and Opus like. Thinks a lot. Goes on long self testing loops/talking with itself, but need more time with it. It seems like a good model for agentic things.

David Dalcu

122,199 просмотров • 27 дней назад

qwen 3.8 max vs deepseek v4 flash 0731 vs kimi k3 vs gpt 5.6 sol – on rubik's cube and chess four frontier models built a rubik's cube stand and solved it, then built a chess board and played claude opus 5 on it the setup: Nous Research's hermes agent cli on OpenRouter tasks: 1. cube – build a 3d rubik's cube with a cli and a Three.js viewer, then solve an identical scrambled position on your own stand 2. chess – build a 3d chess stand, then play white against claude opus 5 as black, live, one move at a time. no engine, no solver, no opening book on either side. stockfish depth 14 grades every chess ply afterwards; neither player sees the score models: DeepSeek v4 flash 0731, OpenAI gpt-5.6 sol, Kimi.ai kimi k3, Qwen qwen 3.8 max gpt-5.6 sol and deepseek v4 flash solved their cubes – sol in 24 moves and seventeen seconds, deepseek in 32. qwen and kimi never got there, giving up at 96 and 207 moves then all four built chess stands and played white against claude opus 5 on them, and all four resigned: deepseek on move 13, sol on 19, kimi on 21, qwen holding out longest at 29 - build time, both stands #1 gpt-5.6 sol – 16m 43s #2 deepseek v4 flash – 97m 39s #3 kimi k3 – 166m 09s #4 qwen 3.8 max – 215m 08s - build attempts before a working stand #1 gpt-5.6 sol – 3 #2 qwen 3.8 max – 4 #3 kimi k3 – 4 #4 deepseek v4 flash – 5 - total tokens #1 gpt-5.6 sol – 6,713,754 #2 qwen 3.8 max – 17,272,507 #3 kimi k3 – 22,427,504 #4 deepseek v4 flash – 27,417,442 - total price #1 deepseek v4 flash – $0.557 #2 gpt-5.6 sol – $6.319 #3 qwen 3.8 max – $10.270 #4 kimi k3 – $16.667 observations: • deepseek v4 flash is the cheapest model here by a margin nobody else is near, and it got there while being the least efficient of the four. it burned 27.4m tokens – more than anyone, 5m more than kimi – and still finished both benchmarks for $0.557. that is $0.02 per million tokens against kimi's $0.74. it also needed the most passes to produce working stands, five, and that did not matter: all five deepseek passes together cost a thirtieth of kimi's two • so what deepseek cannot do is get it right the first time. what it can do is get it right the fifth time, for half a dollar. that is a different thing to be buying – not a good first draft, but the option to keep asking • gpt-5.6 sol is the opposite profile and the strongest of the four on pure efficiency. 16m 43s to build both stands, 6.7m tokens, three passes – under 40% of the next lowest token count and a quarter of deepseek's, on an eighth of qwen's clock. it also solved the cube fastest of anyone, 24 moves in seventeen seconds. sol is what you reach for when you want the answer now and can absorb $0.94 per million • sol's weakness is in what it does not check. its chess viewer deleted the capturing piece instead of the captured one, so pieces disappeared off the board mid-game – a defect the fifty-cent deepseek stand did not have. fast and terse turns out to be the same dial as fast and unverified • qwen 3.8 max is not the cheap open-weights option it gets treated as. $10.270 across the two benchmarks, second most expensive of the four, 18x deepseek, and by a distance the slowest – 215 minutes of build time, nearly thirteen times sol's. what the money buys is judgment: it played eighteen moves without a single error worth a hundredth of a pawn, then made exactly one bad move in the whole game, and averaged 44.6 centipawns lost across the longest game any of the four managed. it also could not solve a rubik's cube in 96 tries • kimi k3 is the one line with no reading that flatters it. most expensive at $16.667, last on the cube at 207 moves, last at chess at 478 centipawns lost per move. it is also the model that verified hardest – on the cube it wrote its own integrity check instead of trusting its output. that makes the result worse rather than better: the checking was real, and the reasoning underneath it still was not follow thehype. for 24/7 ai news, analysis and breakdowns
1:03

qwen 3.8 max vs deepseek v4 flash 0731 vs kimi k3 vs gpt 5.6 sol – on rubik's cube and chess four frontier models built a rubik's cube stand and solved it, then built a chess board and played claude opus 5 on it the setup: Nous Research's hermes agent cli on OpenRouter tasks: 1. cube – build a 3d rubik's cube with a cli and a Three.js viewer, then solve an identical scrambled position on your own stand 2. chess – build a 3d chess stand, then play white against claude opus 5 as black, live, one move at a time. no engine, no solver, no opening book on either side. stockfish depth 14 grades every chess ply afterwards; neither player sees the score models: DeepSeek v4 flash 0731, OpenAI gpt-5.6 sol, Kimi.ai kimi k3, Qwen qwen 3.8 max gpt-5.6 sol and deepseek v4 flash solved their cubes – sol in 24 moves and seventeen seconds, deepseek in 32. qwen and kimi never got there, giving up at 96 and 207 moves then all four built chess stands and played white against claude opus 5 on them, and all four resigned: deepseek on move 13, sol on 19, kimi on 21, qwen holding out longest at 29 - build time, both stands #1 gpt-5.6 sol – 16m 43s #2 deepseek v4 flash – 97m 39s #3 kimi k3 – 166m 09s #4 qwen 3.8 max – 215m 08s - build attempts before a working stand #1 gpt-5.6 sol – 3 #2 qwen 3.8 max – 4 #3 kimi k3 – 4 #4 deepseek v4 flash – 5 - total tokens #1 gpt-5.6 sol – 6,713,754 #2 qwen 3.8 max – 17,272,507 #3 kimi k3 – 22,427,504 #4 deepseek v4 flash – 27,417,442 - total price #1 deepseek v4 flash – $0.557 #2 gpt-5.6 sol – $6.319 #3 qwen 3.8 max – $10.270 #4 kimi k3 – $16.667 observations: • deepseek v4 flash is the cheapest model here by a margin nobody else is near, and it got there while being the least efficient of the four. it burned 27.4m tokens – more than anyone, 5m more than kimi – and still finished both benchmarks for $0.557. that is $0.02 per million tokens against kimi's $0.74. it also needed the most passes to produce working stands, five, and that did not matter: all five deepseek passes together cost a thirtieth of kimi's two • so what deepseek cannot do is get it right the first time. what it can do is get it right the fifth time, for half a dollar. that is a different thing to be buying – not a good first draft, but the option to keep asking • gpt-5.6 sol is the opposite profile and the strongest of the four on pure efficiency. 16m 43s to build both stands, 6.7m tokens, three passes – under 40% of the next lowest token count and a quarter of deepseek's, on an eighth of qwen's clock. it also solved the cube fastest of anyone, 24 moves in seventeen seconds. sol is what you reach for when you want the answer now and can absorb $0.94 per million • sol's weakness is in what it does not check. its chess viewer deleted the capturing piece instead of the captured one, so pieces disappeared off the board mid-game – a defect the fifty-cent deepseek stand did not have. fast and terse turns out to be the same dial as fast and unverified • qwen 3.8 max is not the cheap open-weights option it gets treated as. $10.270 across the two benchmarks, second most expensive of the four, 18x deepseek, and by a distance the slowest – 215 minutes of build time, nearly thirteen times sol's. what the money buys is judgment: it played eighteen moves without a single error worth a hundredth of a pawn, then made exactly one bad move in the whole game, and averaged 44.6 centipawns lost across the longest game any of the four managed. it also could not solve a rubik's cube in 96 tries • kimi k3 is the one line with no reading that flatters it. most expensive at $16.667, last on the cube at 207 moves, last at chess at 478 centipawns lost per move. it is also the model that verified hardest – on the cube it wrote its own integrity check instead of trusting its output. that makes the result worse rather than better: the checking was real, and the reasoning underneath it still was not follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

84,777 просмотров • 1 месяц назад