Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Qwen3.8-Flash-Next just got a massive speed boost on my old hardware. This guy started with roughly 11 tok/s. After tuning the inference stack, I’m now getting around 24–28 tok/s. That’s roughly a 2× speedup without changing the GPU. The setup: → RTX 3090 24GB → 64GB DDR4 system RAM... → Qwen3.8-Flash-Next → Cafe-llama.cpp The interesting part is that Cafe-llama.cpp isn’t just another llama.cpp fork. The optimizations are making a real difference for this particular architecture, especially when you’re dealing with a model that relies heavily on system memory and offloading. Going from 11 tok/s to nearly 30 tok/s completely changes how practical the model feels for local use. And now I’m curious what happens with DDR5. If memory bandwidth becomes another bottleneck, a faster RAM subsystem could squeeze even more performance out of the same 3090. This is why local AI optimization is getting so interesting. Sometimes you don’t need a new GPU. You just need a better inference stack. 11 tok/s → 24–28 tok/s.show more

FHILY👑

9,111 subscribers

23,243 views • 13 days ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

13 Comments

Peasant Smith13 days ago

I tried that fork on my hardware and -unfortunately- it didn't move a needle from what I was getting before. My hardware is probably the limiting factor here. I was already getting 14.1t/s decode on 3x 3060 12gb and ddr4 2400 quad channel. With this fork I got pretty much 14t/s

Stanislaw Weston13 days ago

This is a very good fork of llama.cpp. It has many useful features for MoE models, especially Qwen 3.8 Flash Next. BUT MTP actually slows things down by 1.5x. So while I use this fork, I keep MTP turned off.

Mohammad Zoheb13 days ago

I am using on server, with 158GB RAM + RTX3090 but I only allow 12GB VRAM since I have other apps to use it. Getting 17 to 18 token per second with original llama.cpp. I tested the cafe llama.cpp and made no difference. It's just basic optimizations, It is just another llama.cpp

Abdarahmane Traore13 days ago

2x from tuning alone is the real story with this model, the kernels are young: rebuilding llama.cpp at the merged PR head nearly doubled decode for us too. Next big one for your card: MTP spec decode, once GGUFs get re-exported with the head (PR #27836). Recipes:

LocalAIMaxxer // Niko13 days ago

Could I do this with 5080 + 96gb ram and ssd? How much sits on the gpu?

Diego Garcia13 days ago

Hey mate Mind to share the recipe you used? Thank you!

Alois Vaclav13 days ago

Stop this bullshit, 30tps cunt be your daily driver...nice experiments, but thats all. Cunt call it "locall AI" when it can be used as Codex / Claude..makes no sense.

why13 days ago

Which part of the stack made the biggest difference: quantization, batching, kernel settings, or CPU offload?

LinuxVersion13 days ago

any way to quantize for use with 8GB vram?

De doncker kenny13 days ago

weird, i am on 17 tok/s on 8 gb vram now.

Zima13 days ago

If I want to build this system. What’s a brief shopping list?

Binsen Huang13 days ago

我的用2x5090 32GB，drr5 96g出现了预充填时间过长和解码速度过于缓慢，模型无法正常使用的情况，

Idriss13 days ago

2x on the same 3090 usually means you stopped spilling into system RAM, not that the kernels got faster. Worth reporting tok/s at 8k and 32k context too, since that gap is where the tuning actually shows up.

Related Videos

125B on a single RTX 4090 at 25 tok/s is actually insane Qwen3.8-Flash-Next + MTP is hitting: 80K context 25.35 tok/s decode 471 tok/s prefill On one consumer GPU. Yes, it’s an MoE, so you’re not running all 125B parameters for every token. But still, getting a model this large to run this fast on a 4090 is wild. MTP is doing a lot of the heavy lifting here too. Instead of generating one token at a time, it can speculate on multiple tokens and verify them together. And 25 tok/s is the important number.

125B on a single RTX 4090 at 25 tok/s is actually insane Qwen3.8-Flash-Next + MTP is hitting: 80K context 25.35 tok/s decode 471 tok/s prefill On one consumer GPU. Yes, it’s an MoE, so you’re not running all 125B parameters for every token. But still, getting a model this large to run this fast on a 4090 is wild. MTP is doing a lot of the heavy lifting here too. Instead of generating one token at a time, it can speculate on multiple tokens and verify them together. And 25 tok/s is the important number.

FHILY👑

38,386 views • 11 days ago

🤯 A localmaxxer hit ~381 tok/s on a SINGLE RTX 3090 with Qwen3.8-27B. This developer has turned a 24GB RTX 3090 into a monster Qwen inference box - w/some creativity. Four days ago 👉 ⚡ ~82 tok/s single-user Then 👉 ⚡ ~114 tok/s with optimized MTP ⚡ ~138 tok/s with DFlash2 + lookup drafting Now 👉 🔥 ~381 tok/s on ONE request How? The recipe combines ... 🧠 Qwen3.8-27B 🎮 1× RTX 3090 24GB @ 250W ⚙️ heavily optimized vLLM ⚡ DFlash2 speculative decoding 🔎 lookup-augmented drafting 📚 prefix caching 🧮 16-token verification blocks 💾 quantized KV / heads / activations DFlash2 normally proposes 7 tokens. The developer realized the verification block doesn't have to stop there. If Qwen is answering from a document already sitting in the prompt, the system can fill the remaining draft positions using tokens found directly in that context. 🎯 So the target model can verify 16 tokens at once. On a ~25K-token document reproduction task: Previous DFlash2 👉 ~260 tok/s Longer verification + context lookup 👉 🔥 ~382 tok/s Acceptance: 🤯 15 of 16 tokens per verification step That is where the crazy number comes from. ⚠️ On ordinary real-world chat prompts, the same setup is around ~133 tok/s Still extremely fast for a dense 27B model on an RTX 3090. The 381 tok/s mode shines when the answer largely comes from material already in context so these are best use cases 📚 RAG / document Q&A 💻 Coding assistants applying edits 📝 Quoting or rewriting documents 🔎 Extracting information from long prompts And another optimization 👉 With prefix caching, a second question against the same 25K-token document reportedly goes from: 🐌 22.4 sec TTFT → ⚡ 0.56 sec TTFT Because the model doesn't need to process the whole document again. 🎯 It's specifically a mode for RAG front ends and coding agents. Follow iamMess on Reddit or syv-ai on GitHub 🔗 Reddit: r/LocalLLaMA/comments/1vtup5s/ 🔗 GitHub: /syv-ai/qwen38-27b-rtx3090

🤯 A localmaxxer hit ~381 tok/s on a SINGLE RTX 3090 with Qwen3.8-27B. This developer has turned a 24GB RTX 3090 into a monster Qwen inference box - w/some creativity. Four days ago 👉 ⚡ ~82 tok/s single-user Then 👉 ⚡ ~114 tok/s with optimized MTP ⚡ ~138 tok/s with DFlash2 + lookup drafting Now 👉 🔥 ~381 tok/s on ONE request How? The recipe combines ... 🧠 Qwen3.8-27B 🎮 1× RTX 3090 24GB @ 250W ⚙️ heavily optimized vLLM ⚡ DFlash2 speculative decoding 🔎 lookup-augmented drafting 📚 prefix caching 🧮 16-token verification blocks 💾 quantized KV / heads / activations DFlash2 normally proposes 7 tokens. The developer realized the verification block doesn't have to stop there. If Qwen is answering from a document already sitting in the prompt, the system can fill the remaining draft positions using tokens found directly in that context. 🎯 So the target model can verify 16 tokens at once. On a ~25K-token document reproduction task: Previous DFlash2 👉 ~260 tok/s Longer verification + context lookup 👉 🔥 ~382 tok/s Acceptance: 🤯 15 of 16 tokens per verification step That is where the crazy number comes from. ⚠️ On ordinary real-world chat prompts, the same setup is around ~133 tok/s Still extremely fast for a dense 27B model on an RTX 3090. The 381 tok/s mode shines when the answer largely comes from material already in context so these are best use cases 📚 RAG / document Q&A 💻 Coding assistants applying edits 📝 Quoting or rewriting documents 🔎 Extracting information from long prompts And another optimization 👉 With prefix caching, a second question against the same 25K-token document reportedly goes from: 🐌 22.4 sec TTFT → ⚡ 0.56 sec TTFT Because the model doesn't need to process the whole document again. 🎯 It's specifically a mode for RAG front ends and coding agents. Follow iamMess on Reddit or syv-ai on GitHub 🔗 Reddit: r/LocalLLaMA/comments/1vtup5s/ 🔗 GitHub: /syv-ai/qwen38-27b-rtx3090

David Hendrickson

113,342 views • 24 days ago

"Why are you benchmarking DGX Spark? It's a training box." Yeah. Low bandwidth, but 128GB of unified memory is just sitting there. Plenty of room to optimize. DGX Spark + Qwen3.6 27B. Four backend/quant combos: 🔴 llama.cpp + UD_Q4_K_XL > 11.0 tok/s (baseline), TTFT 297ms 🟢 llama.cpp + DFlash > 20.4 tok/s (peaks at 97 tok/s), TTFT 320ms 🟡 vLLM FP8 + MTP > 13.1 tok/s, TTFT 540ms 🟣 vLLM NVFP4 + MTP > 24.2 tok/s, TTFT 376ms NVFP4+MTP is the winner for me, rock stable around 24 tok/s, no wild swings. DFlash is the wildcard: massive peaks, but fluctuates a lot. FP8+MTP barely beats baseline, and it's FP8. Love my Spark.

"Why are you benchmarking DGX Spark? It's a training box." Yeah. Low bandwidth, but 128GB of unified memory is just sitting there. Plenty of room to optimize. DGX Spark + Qwen3.6 27B. Four backend/quant combos: 🔴 llama.cpp + UD_Q4_K_XL > 11.0 tok/s (baseline), TTFT 297ms 🟢 llama.cpp + DFlash > 20.4 tok/s (peaks at 97 tok/s), TTFT 320ms 🟡 vLLM FP8 + MTP > 13.1 tok/s, TTFT 540ms 🟣 vLLM NVFP4 + MTP > 24.2 tok/s, TTFT 376ms NVFP4+MTP is the winner for me, rock stable around 24 tok/s, no wild swings. DFlash is the wildcard: massive peaks, but fluctuates a lot. FP8+MTP barely beats baseline, and it's FP8. Love my Spark.

stevibe

40,223 views • 4 months ago

Qwen3.6 35B-A3B dropped yesterday, so I ran it on 4 GPUs to see how it performs: 🟣 RTX 3090 — 49.78 tok/s, TTFT 852ms 🟡 RTX 4090 — 118.93 tok/s, TTFT 686ms 🟢 RTX 5090 — 160.37 tok/s, TTFT 409ms 🔵 DGX Spark — 59.98 tok/s, TTFT 228ms I went with ollama as the backend because honestly, it's the easiest way for most people to get started. One command, model pulled, done. I used Q4_K_M (24GB) across all four cards. The reason is the 3090 and 4090 don't support NVFP4 (only the 5090 and DGX Spark could use it). Keeping the same quant everywhere felt like the fairest way to compare. And yes, you can absolutely squeeze more performance out of every card with vLLM, SGLang, or TensorRT-LLM. But that's not what this test is about. This is just the out-of-the-box experience for folks who own a GPU and want to try the new model tonight.

Qwen3.6 35B-A3B dropped yesterday, so I ran it on 4 GPUs to see how it performs: 🟣 RTX 3090 — 49.78 tok/s, TTFT 852ms 🟡 RTX 4090 — 118.93 tok/s, TTFT 686ms 🟢 RTX 5090 — 160.37 tok/s, TTFT 409ms 🔵 DGX Spark — 59.98 tok/s, TTFT 228ms I went with ollama as the backend because honestly, it's the easiest way for most people to get started. One command, model pulled, done. I used Q4_K_M (24GB) across all four cards. The reason is the 3090 and 4090 don't support NVFP4 (only the 5090 and DGX Spark could use it). Keeping the same quant everywhere felt like the fairest way to compare. And yes, you can absolutely squeeze more performance out of every card with vLLM, SGLang, or TensorRT-LLM. But that's not what this test is about. This is just the out-of-the-box experience for folks who own a GPU and want to try the new model tonight.

stevibe

403,049 views • 4 months ago

I know some bros want to see what happens when we push the value even further, so here we go: ⚪️ -ncmoe 25 → 41.8 tok/s RAM: 12GB & VRAM: 6.9GB 🔴 -ncmoe 23 → 43.8 tok/s RAM: 12.2GB & VRAM: 7.4GB 🟢 -ncmoe 21 → 38.6 tok/s RAM: 12.4GB & VRAM: 7.8GB 🔵 -ncmoe 19 → 19.8 tok/s RAM: 13.8GB & VRAM: 7.8GB As you can see, there's a sweet spot with the VRAM usage. Play around and monitor to find the right value for your setup, you can use llama.cpp web ui to monitor speeds easily Sweet spot seems to be 25-23 for 8GB VRAM ✅ +40tok/s for Qwen3.6 35B with 64k q8_0 context on a 8GB card is very impressive just by using base llama.cpp, and we didn't even try Turboquant, MTP or Dflash yet! I'll focus on these next 👀 (Server flags and setup in the quoted tweet)

I know some bros want to see what happens when we push the value even further, so here we go: ⚪️ -ncmoe 25 → 41.8 tok/s RAM: 12GB & VRAM: 6.9GB 🔴 -ncmoe 23 → 43.8 tok/s RAM: 12.2GB & VRAM: 7.4GB 🟢 -ncmoe 21 → 38.6 tok/s RAM: 12.4GB & VRAM: 7.8GB 🔵 -ncmoe 19 → 19.8 tok/s RAM: 13.8GB & VRAM: 7.8GB As you can see, there's a sweet spot with the VRAM usage. Play around and monitor to find the right value for your setup, you can use llama.cpp web ui to monitor speeds easily Sweet spot seems to be 25-23 for 8GB VRAM ✅ +40tok/s for Qwen3.6 35B with 64k q8_0 context on a 8GB card is very impressive just by using base llama.cpp, and we didn't even try Turboquant, MTP or Dflash yet! I'll focus on these next 👀 (Server flags and setup in the quoted tweet)

left curve dev

18,049 views • 4 months ago

Qwen3.6 27B landed yesterday, so I ran it on 4 setups side-by-side to see how they stack up: 🔴 RTX 4090 — 45.59 tok/s, TTFT 525ms 🟢 RTX 5090 — 51.83 tok/s, TTFT 752ms ⚫️ M2 Ultra — 22.30 tok/s, TTFT 216ms 🟣 DGX Spark — 11.08 tok/s, TTFT 319ms This is a standard test: no tuning, just the out-of-the-box experience. For the NVIDIA cards I used llama.cpp with Unsloth's UD-Q4_K_XL quant. For the M2 Ultra I used MLX with Unsloth's UD-MLX-4bit quant, since MLX is the native path on Apple Silicon. Please consider this as the baseline, you can definitely squeeze more out of every one of these with fine-tuned settings.

Qwen3.6 27B landed yesterday, so I ran it on 4 setups side-by-side to see how they stack up: 🔴 RTX 4090 — 45.59 tok/s, TTFT 525ms 🟢 RTX 5090 — 51.83 tok/s, TTFT 752ms ⚫️ M2 Ultra — 22.30 tok/s, TTFT 216ms 🟣 DGX Spark — 11.08 tok/s, TTFT 319ms This is a standard test: no tuning, just the out-of-the-box experience. For the NVIDIA cards I used llama.cpp with Unsloth's UD-Q4_K_XL quant. For the M2 Ultra I used MLX with Unsloth's UD-MLX-4bit quant, since MLX is the native path on Apple Silicon. Please consider this as the baseline, you can definitely squeeze more out of every one of these with fine-tuned settings.

stevibe

104,692 views • 4 months ago

i spent 3 hours finding the sweet spot for hermes 4.3 36B on a single RTX 3090. saving you the trouble, anon. the model is 21.8GB at Q4_K_M. that leaves 2.2GB free on 24GB VRAM. not much room for KV cache. here's what actually happened: 4K: 35.3 tok/s 8K: 35.2 tok/s 16K: 34.8 tok/s 32K: 34.6 tok/s 64K: 6.4 tok/s 128K: 1.9 tok/s flat from 4K to 32K. then it falls off a cliff at 64K. the trick is quantized KV cache. without it you OOM at 16K. with quantized KV cache you get 32K at full speed. all 64 layers on GPU. at 64K something weird happens. ngl 99 (all layers on GPU) = 3.96 tok/s. the KV cache silently spills to CPU. drop to ngl 55 and speed jumps to 6.37. drop to ngl 48 and it gets worse again (3.46). there's an offload sweet spot where you free just enough VRAM for the cache without losing too much compute to PCIe transfers. 128K works at ngl 32 but you're at 1.95 tok/s. half the model on CPU. usable for batch work, not for interactive. the sweet spot command: llama-server -m hermes-4.3-36b-Q4_K_M.gguf -ngl 99 -c 32768 --cache-type-k q4_0 --cache-type-v q4_0 32K context. 34.6 tok/s. all on GPU. this is where dense 36B lives on 24GB. for comparison, qwen 3.5 (35B MoE, 3B active) holds 112 tok/s from 4K all the way to 262K on the same GPU. no speed drop. same total params, completely different architecture. hybrid linear attention means flat context scaling. dense pays for every token in the KV cache. code and quality comparison coming next. fast vs slow generation side by side in the videos below.

i spent 3 hours finding the sweet spot for hermes 4.3 36B on a single RTX 3090. saving you the trouble, anon. the model is 21.8GB at Q4_K_M. that leaves 2.2GB free on 24GB VRAM. not much room for KV cache. here's what actually happened: 4K: 35.3 tok/s 8K: 35.2 tok/s 16K: 34.8 tok/s 32K: 34.6 tok/s 64K: 6.4 tok/s 128K: 1.9 tok/s flat from 4K to 32K. then it falls off a cliff at 64K. the trick is quantized KV cache. without it you OOM at 16K. with quantized KV cache you get 32K at full speed. all 64 layers on GPU. at 64K something weird happens. ngl 99 (all layers on GPU) = 3.96 tok/s. the KV cache silently spills to CPU. drop to ngl 55 and speed jumps to 6.37. drop to ngl 48 and it gets worse again (3.46). there's an offload sweet spot where you free just enough VRAM for the cache without losing too much compute to PCIe transfers. 128K works at ngl 32 but you're at 1.95 tok/s. half the model on CPU. usable for batch work, not for interactive. the sweet spot command: llama-server -m hermes-4.3-36b-Q4_K_M.gguf -ngl 99 -c 32768 --cache-type-k q4_0 --cache-type-v q4_0 32K context. 34.6 tok/s. all on GPU. this is where dense 36B lives on 24GB. for comparison, qwen 3.5 (35B MoE, 3B active) holds 112 tok/s from 4K all the way to 262K on the same GPU. no speed drop. same total params, completely different architecture. hybrid linear attention means flat context scaling. dense pays for every token in the KV cache. code and quality comparison coming next. fast vs slow generation side by side in the videos below.

Sudo su

52,268 views • 6 months ago

Anyone with 8GB or 12GB VRAM setups needs to understand that "-ncmoe" is the key flag to boost performance on llama.cpp Here are my results for Qwen3.6 35B A3B, with 64k q8_0 context on a 8GB RTX 3070Ti: ⚪️ no flag → 8.7 tok/s RAM: 13.6GB & VRAM: 7.8GB 🔴 -ncmoe 35 → 27.5 tok/s RAM: 12.1GB & VRAM: 4.3GB 🟢 -ncmoe 30 → 32.5 tok/s RAM: 12GB & VRAM: 5.6GB 🔵 -ncmoe 25 → 40.9 tok/s RAM: 12GB & VRAM: 6.9GB Please note the ram and vram usage you see are total usage of a windows pc, with the model running. My friend's setup: 8GB VRAM and 16GB RAM. You can boost performance by switching to Linux, just something to keep in mind. Basically, this flag keeps the MoE experts in the first X layers on your CPU + RAM, instead of eating all your VRAM straight away. This is a smart hybrid offload way that lets you run bigger models without OOM while keeping the rest on your GPU for speed. As we can see on the data, there's a sweet spot. When we lower it from 35 to 25, speed bumps +50% because there are more layers on your GPU (look at the VRAM usage). The key here is to play around with the number and fit as much as possible on your VRAM, goal is to have 1GB/800MB headroom to avoid stress. ↓ server flags below

Anyone with 8GB or 12GB VRAM setups needs to understand that "-ncmoe" is the key flag to boost performance on llama.cpp Here are my results for Qwen3.6 35B A3B, with 64k q8_0 context on a 8GB RTX 3070Ti: ⚪️ no flag → 8.7 tok/s RAM: 13.6GB & VRAM: 7.8GB 🔴 -ncmoe 35 → 27.5 tok/s RAM: 12.1GB & VRAM: 4.3GB 🟢 -ncmoe 30 → 32.5 tok/s RAM: 12GB & VRAM: 5.6GB 🔵 -ncmoe 25 → 40.9 tok/s RAM: 12GB & VRAM: 6.9GB Please note the ram and vram usage you see are total usage of a windows pc, with the model running. My friend's setup: 8GB VRAM and 16GB RAM. You can boost performance by switching to Linux, just something to keep in mind. Basically, this flag keeps the MoE experts in the first X layers on your CPU + RAM, instead of eating all your VRAM straight away. This is a smart hybrid offload way that lets you run bigger models without OOM while keeping the rest on your GPU for speed. As we can see on the data, there's a sweet spot. When we lower it from 35 to 25, speed bumps +50% because there are more layers on your GPU (look at the VRAM usage). The key here is to play around with the number and fit as much as possible on your VRAM, goal is to have 1GB/800MB headroom to avoid stress. ↓ server flags below

left curve dev

166,478 views • 4 months ago

Qwen3.8-Flash-Next now fits on a 48gb macbook air Got q4 running on MLX using only 39gb of memory and decoding at 28 tok/s 🔥 Also q8 running on just 50gb of memory I stacked two tricks so far: 1. REAP pruning cut the expert pool from 68gb to 35gb 2. the 51b n-gram table now lives on the ssd instead of ram My patch memmaps the shards and reads ~100 bytes per lookup straight off disk, dequantized on cpu and verified bit-exact against the gpu path Qwen3.8-Flash-Next at 288 experts with its PLE table on nvme loads in 6 seconds, 39gb resident, 28 tok/s decode, 600 tok/s prefill. Logits identical to the ram-resident version in every way 97gb stock Q4 down to 68gb if pruned to 288 experts down to 39gb if pruned to 288 experts with the table on disk humaneval scoring 91.5% with 288 experts vs. 93.9% for stock I'm not done. next: 1. MTP drafter for speculative decode 2. n-gram lookup drafter both are built, benchmarking them now

Qwen3.8-Flash-Next now fits on a 48gb macbook air Got q4 running on MLX using only 39gb of memory and decoding at 28 tok/s 🔥 Also q8 running on just 50gb of memory I stacked two tricks so far: 1. REAP pruning cut the expert pool from 68gb to 35gb 2. the 51b n-gram table now lives on the ssd instead of ram My patch memmaps the shards and reads ~100 bytes per lookup straight off disk, dequantized on cpu and verified bit-exact against the gpu path Qwen3.8-Flash-Next at 288 experts with its PLE table on nvme loads in 6 seconds, 39gb resident, 28 tok/s decode, 600 tok/s prefill. Logits identical to the ram-resident version in every way 97gb stock Q4 down to 68gb if pruned to 288 experts down to 39gb if pruned to 288 experts with the table on disk humaneval scoring 91.5% with 288 experts vs. 93.9% for stock I'm not done. next: 1. MTP drafter for speculative decode 2. n-gram lookup drafter both are built, benchmarking them now

Eyal Toledano

10,440 views • 17 days ago

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.

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

71,112 views • 5 months ago

Someone bought a Mac Mini for local AI. 64GB of unified memory seemed like the obvious choice. More memory. Bigger models. No VRAM limits. Then they put it against an RTX 3060. And things got weird. The Mac Mini M4 Pro pushed Qwen 2.5 at 38 tok/s. The old RTX 3060 with just 12GB VRAM hit 52 tok/s. That’s the interesting part about local AI. Apple and NVIDIA are solving different problems. Apple gives you a huge pool of memory. NVIDIA gives you dedicated compute and CUDA. One helps you fit bigger models. The other can run smaller models faster. Memory gets the model in. Compute gets the tokens out.

Someone bought a Mac Mini for local AI. 64GB of unified memory seemed like the obvious choice. More memory. Bigger models. No VRAM limits. Then they put it against an RTX 3060. And things got weird. The Mac Mini M4 Pro pushed Qwen 2.5 at 38 tok/s. The old RTX 3060 with just 12GB VRAM hit 52 tok/s. That’s the interesting part about local AI. Apple and NVIDIA are solving different problems. Apple gives you a huge pool of memory. NVIDIA gives you dedicated compute and CUDA. One helps you fit bigger models. The other can run smaller models faster. Memory gets the model in. Compute gets the tokens out.

KyzoroX

26,269 views • 28 days ago

$RTX 3090 was never designed to be a server GPU. Yet someone just built a 4-GPU AI workstation around four of them. 4× RTX 3090 96GB VRAM 256GB ECC RAM AMD EPYC 7K62 2000W PSU 11 fans The interesting part isn't the performance. It's the memory. Four 24GB cards can be found for a fraction of what you'd pay for modern datacenter GPUs with similar total VRAM. The catch is obvious: 96GB doesn't behave like one 96GB GPU. You still have to deal with GPU-to-GPU communication, power and heat. But for local AI, this is exactly why the 3090 refuses to disappear. Would you build this today?$

RTX 3090 was never designed to be a server GPU. Yet someone just built a 4-GPU AI workstation around four of them. 4× RTX 3090 96GB VRAM 256GB ECC RAM AMD EPYC 7K62 2000W PSU 11 fans The interesting part isn't the performance. It's the memory. Four 24GB cards can be found for a fraction of what you'd pay for modern datacenter GPUs with similar total VRAM. The catch is obvious: 96GB doesn't behave like one 96GB GPU. You still have to deal with GPU-to-GPU communication, power and heat. But for local AI, this is exactly why the 3090 refuses to disappear. Would you build this today?

KyzoroX

36,552 views • 7 days ago

We used to buy more VRAM. Now we’re finding ways around the VRAM limit. This little DDR5 SO-DIMM setup is paired with an old NVIDIA Tesla P40 in the video. The P40 is an old datacenter GPU with 24GB of VRAM. By today’s standards, it’s not fast. But combine that with a machine carrying 128GB of system memory, and suddenly you have a very different local AI box. The GPU handles what it’s good at. The system memory gives the model somewhere to go when VRAM runs out. You obviously pay for that with bandwidth and inference speed. But this is the part I find interesting: Local AI is making “how much memory do I have?” almost as important as “how powerful is my GPU?” We’ve spent years chasing faster GPUs. Maybe the next wave of local AI hardware will be about finding smarter ways to use all the memory we already have. More VRAM is still better. But it’s no longer the only way to run a bigger model.

We used to buy more VRAM. Now we’re finding ways around the VRAM limit. This little DDR5 SO-DIMM setup is paired with an old NVIDIA Tesla P40 in the video. The P40 is an old datacenter GPU with 24GB of VRAM. By today’s standards, it’s not fast. But combine that with a machine carrying 128GB of system memory, and suddenly you have a very different local AI box. The GPU handles what it’s good at. The system memory gives the model somewhere to go when VRAM runs out. You obviously pay for that with bandwidth and inference speed. But this is the part I find interesting: Local AI is making “how much memory do I have?” almost as important as “how powerful is my GPU?” We’ve spent years chasing faster GPUs. Maybe the next wave of local AI hardware will be about finding smarter ways to use all the memory we already have. More VRAM is still better. But it’s no longer the only way to run a bigger model.

KyzoroX

35,554 views • 4 days ago

DFlash makes Qwen 2.2x faster with no quality loss! We ran the same Qwen3.6-27B locally three ways on one RTX 6000: baseline, MTP, DFlash. The tasks only differ in one thing - how predictable the next word is: quicksort, describe a file in JSON, a logic puzzle, a sci-fi story. Outputs: Baseline: 44 tok/s · 1.00x MTP: 65 tok/s · 1.45x · 71% accepted DFlash: 98 tok/s · 2.20x · 30% accepted Baseline writes one token per step. MTP works inside the model itself and guesses 3 tokens ahead. DFlash is a separate small model that writes 15 tokens at once, and the big model only checks them. In JSON the same words repeat all the time, so most guesses were right: 152 tok/s, 3.4x speedup. In the story 9 guesses out of 10 were wrong. DFlash did all that extra work for nothing and became slower than baseline: 42 vs 44 tok/s. MTP guesses only 3 tokens, so a wrong guess costs very little: 46 tok/s and the win in that round. The output is identical in all three modes - DFlash is the pick for tasks with predictable output, like coding, and for chat and creative writing MTP works better. DFlash is now natively integrated into Atomic Chat on llama.cpp - speed up your Qwen models!

DFlash makes Qwen 2.2x faster with no quality loss! We ran the same Qwen3.6-27B locally three ways on one RTX 6000: baseline, MTP, DFlash. The tasks only differ in one thing - how predictable the next word is: quicksort, describe a file in JSON, a logic puzzle, a sci-fi story. Outputs: Baseline: 44 tok/s · 1.00x MTP: 65 tok/s · 1.45x · 71% accepted DFlash: 98 tok/s · 2.20x · 30% accepted Baseline writes one token per step. MTP works inside the model itself and guesses 3 tokens ahead. DFlash is a separate small model that writes 15 tokens at once, and the big model only checks them. In JSON the same words repeat all the time, so most guesses were right: 152 tok/s, 3.4x speedup. In the story 9 guesses out of 10 were wrong. DFlash did all that extra work for nothing and became slower than baseline: 42 vs 44 tok/s. MTP guesses only 3 tokens, so a wrong guess costs very little: 46 tok/s and the win in that round. The output is identical in all three modes - DFlash is the pick for tasks with predictable output, like coding, and for chat and creative writing MTP works better. DFlash is now natively integrated into Atomic Chat on llama.cpp - speed up your Qwen models!

atomic.chat

83,038 views • 2 months ago

Been giving my M3 Ultra Studio some love and tinkering with MLX. Started the afternoon at 42 tok/s. Now: 73 tok/s, and there’s still more on the table. Running Qwen3.8-Flash-Next @ 4-bit locally. Now for the crazy bit: Artificial Analysis Intelligence Index v4.3: GPT-5.6 Sol 39 (Medium) GPT-5.6 Sol 42 (High) Qwen3.8-Flash-Next 42 GPT-5.6 Sol 44 (XHigh) GPT-5.6 Sol 47 (Max) Grok 4.5 39 (High) Claude Sonnet 5 38 (Max) *Qwen score currently estimated by Artificial Analysis. A 4-bit open-weights model running at 73 tok/s on a Mac under my desk... with an estimated intelligence score matching GPT-5.6 Sol at High reasoning. Local AI is getting ridiculous. 🔥 Plus, this model does not say no to a bit of red teaming. Once I am happy with it, I will do a pr for Jun Kim's oMLX :)

Been giving my M3 Ultra Studio some love and tinkering with MLX. Started the afternoon at 42 tok/s. Now: 73 tok/s, and there’s still more on the table. Running Qwen3.8-Flash-Next @ 4-bit locally. Now for the crazy bit: Artificial Analysis Intelligence Index v4.3: GPT-5.6 Sol 39 (Medium) GPT-5.6 Sol 42 (High) Qwen3.8-Flash-Next 42 GPT-5.6 Sol 44 (XHigh) GPT-5.6 Sol 47 (Max) Grok 4.5 39 (High) Claude Sonnet 5 38 (Max) *Qwen score currently estimated by Artificial Analysis. A 4-bit open-weights model running at 73 tok/s on a Mac under my desk... with an estimated intelligence score matching GPT-5.6 Sol at High reasoning. Local AI is getting ridiculous. 🔥 Plus, this model does not say no to a bit of red teaming. Once I am happy with it, I will do a pr for Jun Kim's oMLX :)

Ash Hart

18,827 views • 5 days ago

been playing with hermes agent paired with qwen 3.5 dense 27B on my single 3090 since last night. there is something about this harness that caught me and i think i know what it is. i've now run five qwen configs on consumer hardware: 35B MoE (3B active) -- 112 tok/s flat across 262K context, 1x 3090 27B dense -- 35 tok/s, zero degradation across the same range, 1x 3090 qwopus 27B (opus distilled) -- 35.7 tok/s, same architecture, different brain 80B coder -- 46 tok/s on 2x 3090s, oneshotted a 564 line particle sim 80B coder -- 1.3 tok/s on 1x 3090, bleeding through RAM because it didn't fit but it still ran with same benchmarks. same prompts. same quant where possible. every config is documented. i know these models. and hermes agent is the first harness that feels like it respects that work. tool calls show inline with execution time. nvidia-smi 0.2s. write_file 0.7s. you see exactly what the agent is doing and how long each step takes. no mystery. no black box. no tool call failures so far and i've been pushing it. most agent frameworks feel like you're watching a spinner and hoping. hermes shows the work. that transparency changes how you trust the output. once you use it you see the UX decisions are not accidental. Teknium 🪽 and the nous team built this like engineers who actually use their own tools. 80 skills. 29 tools. persistent memory. context compression. runs clean on a single consumer GPU.

been playing with hermes agent paired with qwen 3.5 dense 27B on my single 3090 since last night. there is something about this harness that caught me and i think i know what it is. i've now run five qwen configs on consumer hardware: 35B MoE (3B active) -- 112 tok/s flat across 262K context, 1x 3090 27B dense -- 35 tok/s, zero degradation across the same range, 1x 3090 qwopus 27B (opus distilled) -- 35.7 tok/s, same architecture, different brain 80B coder -- 46 tok/s on 2x 3090s, oneshotted a 564 line particle sim 80B coder -- 1.3 tok/s on 1x 3090, bleeding through RAM because it didn't fit but it still ran with same benchmarks. same prompts. same quant where possible. every config is documented. i know these models. and hermes agent is the first harness that feels like it respects that work. tool calls show inline with execution time. nvidia-smi 0.2s. write_file 0.7s. you see exactly what the agent is doing and how long each step takes. no mystery. no black box. no tool call failures so far and i've been pushing it. most agent frameworks feel like you're watching a spinner and hoping. hermes shows the work. that transparency changes how you trust the output. once you use it you see the UX decisions are not accidental. Teknium 🪽 and the nous team built this like engineers who actually use their own tools. 80 skills. 29 tools. persistent memory. context compression. runs clean on a single consumer GPU.

Sudo su

100,259 views • 6 months ago

How Fast is Gemma 4 on a MacBook Pro M4? Benchmarking Google's new MoE (26B-A4B) > Model size: 26.1 GiB > Load time: ~4.2s Comparing single request VS > concurrent requests performance > 32k total context, 4 parallel slots single request behavior > TTFT: 5.68s > prompt: 3,701 tokens @ 652 tok/s > decode: 40.08 tok/s sequential (1 request at a time): > avg duration: 20.5s > p99: 22.1s > throughput: 40.11 tok/s > clean finishes: 100% concurrent (4 parallel requests): > aggregate throughput: 47.25 tok/s > total system throughput: 262.27 tok/s > avg duration: 65.1s > p95 latency: 68.8s > req/sec: 0.058 Head-to-Head: Sequential vs Concurrent throughput: > 40.11 tok/s → 47.25 tok/s (+17.8%) > small gain despite 4x parallelism latency per request: > 20.5s → 65.1s (~3.2x slower) > you pay heavily for concurrency system throughput (true utilization): > ~40 tok/s → 262 tok/s (~6.5x total output) > this is where concurrency wins tokens per second (decode ceiling): > ~40 tok/s steady in both modes > hardware-bound, not scheduler-bound TTFT impact: > ~5.7s baseline → buried under queueing in concurrent > “headers waittime” becomes the bottleneck What this actually means? - You don’t get linear scaling from parallel slots - You trade latency for total output - Mac Unified Memory setup is clearly saturating - Bandwidth + Scheduling overhead show up immediately This is exactly why GPUs dominate here Concurrency without killing latency

How Fast is Gemma 4 on a MacBook Pro M4? Benchmarking Google's new MoE (26B-A4B) > Model size: 26.1 GiB > Load time: ~4.2s Comparing single request VS > concurrent requests performance > 32k total context, 4 parallel slots single request behavior > TTFT: 5.68s > prompt: 3,701 tokens @ 652 tok/s > decode: 40.08 tok/s sequential (1 request at a time): > avg duration: 20.5s > p99: 22.1s > throughput: 40.11 tok/s > clean finishes: 100% concurrent (4 parallel requests): > aggregate throughput: 47.25 tok/s > total system throughput: 262.27 tok/s > avg duration: 65.1s > p95 latency: 68.8s > req/sec: 0.058 Head-to-Head: Sequential vs Concurrent throughput: > 40.11 tok/s → 47.25 tok/s (+17.8%) > small gain despite 4x parallelism latency per request: > 20.5s → 65.1s (~3.2x slower) > you pay heavily for concurrency system throughput (true utilization): > ~40 tok/s → 262 tok/s (~6.5x total output) > this is where concurrency wins tokens per second (decode ceiling): > ~40 tok/s steady in both modes > hardware-bound, not scheduler-bound TTFT impact: > ~5.7s baseline → buried under queueing in concurrent > “headers waittime” becomes the bottleneck What this actually means? - You don’t get linear scaling from parallel slots - You trade latency for total output - Mac Unified Memory setup is clearly saturating - Bandwidth + Scheduling overhead show up immediately This is exactly why GPUs dominate here Concurrency without killing latency

Ahmad

88,866 views • 5 months ago

PowerInfer - a high-speed inference engine for deploying LLMs locally. Just came across this super interesting project on speeding up inference. It's not MoE but it's a simple approach that exploits the high locality in LLM inference to design a GPU-CPU hybrid inference engine. Hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons (the majority) are computed on the CPU. This approach significantly reduces GPU memory demands and CPU-GPU data transfer. It achieves an average token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs on a single NVIDIA RTX 4090 GPU. It's on only 18% lower than that achieved by a top-tier server-grade A100 GPU. It also significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy. There is a lot more innovation around inference that's coming fast. Really encouraged by the study on sparse computation to enhance the computational efficiency of LLMs. It's now possible to use PowerInfer with Llama 2 and Faclon 40B. Mistral-7B support is coming soon!

PowerInfer - a high-speed inference engine for deploying LLMs locally. Just came across this super interesting project on speeding up inference. It's not MoE but it's a simple approach that exploits the high locality in LLM inference to design a GPU-CPU hybrid inference engine. Hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons (the majority) are computed on the CPU. This approach significantly reduces GPU memory demands and CPU-GPU data transfer. It achieves an average token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs on a single NVIDIA RTX 4090 GPU. It's on only 18% lower than that achieved by a top-tier server-grade A100 GPU. It also significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy. There is a lot more innovation around inference that's coming fast. Really encouraged by the study on sparse computation to enhance the computational efficiency of LLMs. It's now possible to use PowerInfer with Llama 2 and Faclon 40B. Mistral-7B support is coming soon!

elvis

261,622 views • 2 years ago

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

Reese Chong

53,026 views • 4 months ago

🚨 do you understand what this AMD engineer just proved on camera watch the clip, it’s all there in one frame: llama.cpp slot logs on the left, nvitop showing 79% GPU util holding steady, localhost chat answering a 26k token prompt at 711 tok/s prompt eval, the gemma-4-12B Q4_K_XL gguf running clean. no cloud call, no api key, no censorship layer. for 2 years the rule was: frontier models need the cloud, local is too slow for real work, pay $200/month and wait for tokens. today a $1,800 lunch box runs qwen 3 235B at 11 tok/s with zero data leaving the room, and the same box runs gemma 4 12B with multi token prediction at 14+ tok/s on 30k context with no throttling. $440/month subscription stack: dead. $200/month claude tier: dead. local is too slow excuse: dead. one lunch box. $9/month in electricity. two moats buried in one week. the you need the cloud line wasn’t protecting users it was protecting the subscription.

🚨 do you understand what this AMD engineer just proved on camera watch the clip, it’s all there in one frame: llama.cpp slot logs on the left, nvitop showing 79% GPU util holding steady, localhost chat answering a 26k token prompt at 711 tok/s prompt eval, the gemma-4-12B Q4_K_XL gguf running clean. no cloud call, no api key, no censorship layer. for 2 years the rule was: frontier models need the cloud, local is too slow for real work, pay $200/month and wait for tokens. today a $1,800 lunch box runs qwen 3 235B at 11 tok/s with zero data leaving the room, and the same box runs gemma 4 12B with multi token prediction at 14+ tok/s on 30k context with no throttling. $440/month subscription stack: dead. $200/month claude tier: dead. local is too slow excuse: dead. one lunch box. $9/month in electricity. two moats buried in one week. the you need the cloud line wasn’t protecting users it was protecting the subscription.

AdiiX

22,223 views • 3 months ago