正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Qwen3.5-35B with only 3B active parameters This MoE model runs FASTER than most 7B dense models. Tested on 3 generations of NVIDIA: - 5090: 137 tok/s - 4090: 112 tok/s - 3090: 78 tok/s The surprise? The 4090 <> 5090 gap is only 22%. With a 3B active MoE,... show more

stevibe

27,366 subscribers

69,532 次观看 • 7 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

The RTX 3090 is a 5-year-old GPU and it still runs a 27B model at 20 tok/s I tested Qwen3.5:27b across 3 generations of NVIDIA: 5090 → ~60 tok/s 4090 → ~40 tok/s 3090 → ~20 tok/s Perfectly linear scaling. Double the generation, double the speed.

The RTX 3090 is a 5-year-old GPU and it still runs a 27B model at 20 tok/s I tested Qwen3.5:27b across 3 generations of NVIDIA: 5090 → ~60 tok/s 4090 → ~40 tok/s 3090 → ~20 tok/s Perfectly linear scaling. Double the generation, double the speed.

stevibe

142,493 次观看 • 7 个月前

DGX Spark hit 228ms time-to-first-token on Qwen3.6 35B-A3B. That is 4x faster than an RTX 3090 and 45% ahead of the RTX 5090 Raw throughput still belongs to the 5090 at 160 tok/s. Full numbers, same Ollama backend, Q4_K_M across all four: RTX 3090: 49.78 tok/s, 852ms RTX 4090: 118.93 tok/s, 686ms RTX 5090: 160.37 tok/s, 409ms DGX Spark: 59.98 tok/s, 228ms

DGX Spark hit 228ms time-to-first-token on Qwen3.6 35B-A3B. That is 4x faster than an RTX 3090 and 45% ahead of the RTX 5090 Raw throughput still belongs to the 5090 at 160 tok/s. Full numbers, same Ollama backend, Q4_K_M across all four: RTX 3090: 49.78 tok/s, 852ms RTX 4090: 118.93 tok/s, 686ms RTX 5090: 160.37 tok/s, 409ms DGX Spark: 59.98 tok/s, 228ms

NO1ennn

22,497 次观看 • 2 个月前

Finally got my hands on the big one. Qwen3.5-122B-A10B — 122 billion parameters. Too big for any single consumer GPU. So I rented 4 of each... and then one professional card to see if brute force even matters. - 1x RTX PRO 6000 (96GB): 101.4 tok/s - 4x 5090 (128GB): 87.0 tok/s - 4x 4090 (96GB): 25.1 tok/s - 4x 3090 (96GB): 20.8 tok/s One single $8,500 card beat four RTX 5090s

Finally got my hands on the big one. Qwen3.5-122B-A10B — 122 billion parameters. Too big for any single consumer GPU. So I rented 4 of each... and then one professional card to see if brute force even matters. - 1x RTX PRO 6000 (96GB): 101.4 tok/s - 4x 5090 (128GB): 87.0 tok/s - 4x 4090 (96GB): 25.1 tok/s - 4x 3090 (96GB): 20.8 tok/s One single $8,500 card beat four RTX 5090s

stevibe

197,300 次观看 • 6 个月前

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

403,356 次观看 • 5 个月前

Everyone's comparing the DGX Spark to a 5090 and calling it slow. I think that's the wrong comparison. I ran Qwen3.6 35B-A3B FP8 with the full 262K context window enabled (~96GB RAM) — something gaming GPUs can't really do. Results: 🟢No context: 51.3 tok/s, TTFT 110ms 🟣200K prefill: 34.6 tok/s, TTFT 85s (~2,341 tok/s prefill) Prefill is way faster than a Mac. And 35 tok/s deep into 200K context, on a model this strong, is genuinely usable. The Spark plays a different game.

Everyone's comparing the DGX Spark to a 5090 and calling it slow. I think that's the wrong comparison. I ran Qwen3.6 35B-A3B FP8 with the full 262K context window enabled (~96GB RAM) — something gaming GPUs can't really do. Results: 🟢No context: 51.3 tok/s, TTFT 110ms 🟣200K prefill: 34.6 tok/s, TTFT 85s (~2,341 tok/s prefill) Prefill is way faster than a Mac. And 35 tok/s deep into 200K context, on a model this strong, is genuinely usable. The Spark plays a different game.

stevibe

34,141 次观看 • 5 个月前

4×5090 lost badly in my last Ollama test But with vLLM + Nemotron 3 Super NVFP4, it came back swinging. 4×5090 - TTFT: 0.281s - Prompt eval: 608 tok/s - Generate eval: 110 tok/s PRO 6000 - TTFT: 0.123s - Prompt eval: 1390 tok/s - Generate eval: 81 tok/s So who wins? - PRO 6000 is still faster to respond - But 4×5090 now wins on raw generation speed

4×5090 lost badly in my last Ollama test But with vLLM + Nemotron 3 Super NVFP4, it came back swinging. 4×5090 - TTFT: 0.281s - Prompt eval: 608 tok/s - Generate eval: 110 tok/s PRO 6000 - TTFT: 0.123s - Prompt eval: 1390 tok/s - Generate eval: 81 tok/s So who wins? - PRO 6000 is still faster to respond - But 4×5090 now wins on raw generation speed

stevibe

16,984 次观看 • 6 个月前

128GB in Framework Desktop means you can run giant models locally like AI at Meta's Llama 4 Scout. This is a 109B MoE model with 17B active parameters. At Q6 this runs at >14 tok/s on LM Studio on Fedora 42!

128GB in Framework Desktop means you can run giant models locally like AI at Meta's Llama 4 Scout. This is a 109B MoE model with 17B active parameters. At Q6 this runs at >14 tok/s on LM Studio on Fedora 42!

Framework

30,053 次观看 • 1 年前

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👑

39,979 次观看 • 1 个月前

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 次观看 • 5 个月前

Curious about ollama kimi-k2.6:cloud speed? 3 test runs: > 77.9 tok/s, TTFT 979ms > 114.3 tok/s, TTFT 788ms > 86.3 tok/s, TTFT 1117ms For comparison, OpenRouter stats: > Parasail: 14 tok/s > Moonshot AI: 27 tok/s > NovitaAI: 27 tok/s > Cloudflare: 71 tok/s Obvious caveat: cloud speeds fluctuate with load. Just sharing numbers for the curious.

Curious about ollama kimi-k2.6:cloud speed? 3 test runs: > 77.9 tok/s, TTFT 979ms > 114.3 tok/s, TTFT 788ms > 86.3 tok/s, TTFT 1117ms For comparison, OpenRouter stats: > Parasail: 14 tok/s > Moonshot AI: 27 tok/s > NovitaAI: 27 tok/s > Cloudflare: 71 tok/s Obvious caveat: cloud speeds fluctuate with load. Just sharing numbers for the curious.

stevibe

35,129 次观看 • 5 个月前

Step-3.7-Flash Q4_K_S on DGX Spark (GB10, 128GB): > ~27 tok/s generation > 198B sparse MoE, ~11B active > 256K context, native vision > Agentic / tool-calling / reasoning > Apache 2.0 I added a mobile chat screen on the right showing what 27 tok/s actually feels like streaming on a phone.

Step-3.7-Flash Q4_K_S on DGX Spark (GB10, 128GB): > ~27 tok/s generation > 198B sparse MoE, ~11B active > 256K context, native vision > Agentic / tool-calling / reasoning > Apache 2.0 I added a mobile chat screen on the right showing what 27 tok/s actually feels like streaming on a phone.

stevibe

19,658 次观看 • 4 个月前

GLM-5.2 Nvidia NVFP4 No pruning, no quantising, exact same model card. 110 tok/s single stream. My friend is a genius, holy fudge. Here's the repo, he got full DS4-Flash on a 5090 + DDR5 at 38.5 tok/s This is a breakthrough.

GLM-5.2 Nvidia NVFP4 No pruning, no quantising, exact same model card. 110 tok/s single stream. My friend is a genius, holy fudge. Here's the repo, he got full DS4-Flash on a 5090 + DDR5 at 38.5 tok/s This is a breakthrough.

0xSero

158,188 次观看 • 2 个月前

Qwen3.5:9b reasoning head-to-head: Mac Studio M2 Ultra 64GB: 43.08 tok/s Mac Mini M4 16GB: 13.07 tok/s Qwen

Qwen3.5:9b reasoning head-to-head: Mac Studio M2 Ultra 64GB: 43.08 tok/s Mac Mini M4 16GB: 13.07 tok/s Qwen

stevibe

243,465 次观看 • 7 个月前

Update: Qwen3.5:9b head-to-head (MLX, Apple Silicon optimized) Mac Studio M2 Ultra: 89.74 tok/s Mac Mini M4: 20.82 tok/s MLX basically doubles the speed on both machines.

Update: Qwen3.5:9b head-to-head (MLX, Apple Silicon optimized) Mac Studio M2 Ultra: 89.74 tok/s Mac Mini M4: 20.82 tok/s MLX basically doubles the speed on both machines.

stevibe

134,256 次观看 • 7 个月前

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 次观看 • 7 个月前

Qwen3.6 35B-A3B MTP is a devil performing around 1,000 tok/s on code with RTX 5090 once the prediction gets easier. Been a while I've tried 35B-A3B. Seems I might be sticking around for the fun.

Qwen3.6 35B-A3B MTP is a devil performing around 1,000 tok/s on code with RTX 5090 once the prediction gets easier. Been a while I've tried 35B-A3B. Seems I might be sticking around for the fun.

wd 🇵🇹🔺

49,211 次观看 • 2 个月前

MiniMax M2.7 is 230B params. Can you actually run it at home? I tested Unsloth's UD-IQ3_XXS (80GB) on 4 different rigs: 🟠 4x RTX 4090 (96GB): 71.52 tok/s, TTFT 1045ms 🟢 4x RTX 5090 (128GB): 120.54 tok/s, TTFT 725ms 🟡 1x RTX PRO 6000 (96GB): 118.74 tok/s, TTFT 765ms 🟣 DGX Spark (128GB) — 24.41 tok/s, TTFT 741ms Backend: llama.cpp. Context: 32k. Max tokens: 4096. I went with IQ3_XXS because it's the biggest quant that fits in 96GB VRAM while still leaving safe headroom for 32k context. Same quant across all four rigs, fairest comparison I could run. Now look at rough peak GPU power draw: 🟠 4x4090 → 1,800W peak (450W × 4) 🟢 4x5090 → 2,300W peak (575W × 4) 🟡 RTX PRO 6000 → 600W peak 🟣 DGX Spark → 240W peak (whole system) The RTX PRO 6000 is the quiet winner. One card, 96GB, matching a 4x5090 rig at roughly a quarter of the power and zero multi-GPU headaches. Best tokens-per-watt by a wide margin. DGX Spark is slow on generation but pulls the least power of any rig here, around 240W for the whole system. Prefill-friendly, memory-rich, wall-socket-friendly. And yes, plenty of people cap their cards. Even then, 4x 4090 or 4x 5090 still pulls well over 1,200W from the GPUs alone.

MiniMax M2.7 is 230B params. Can you actually run it at home? I tested Unsloth's UD-IQ3_XXS (80GB) on 4 different rigs: 🟠 4x RTX 4090 (96GB): 71.52 tok/s, TTFT 1045ms 🟢 4x RTX 5090 (128GB): 120.54 tok/s, TTFT 725ms 🟡 1x RTX PRO 6000 (96GB): 118.74 tok/s, TTFT 765ms 🟣 DGX Spark (128GB) — 24.41 tok/s, TTFT 741ms Backend: llama.cpp. Context: 32k. Max tokens: 4096. I went with IQ3_XXS because it's the biggest quant that fits in 96GB VRAM while still leaving safe headroom for 32k context. Same quant across all four rigs, fairest comparison I could run. Now look at rough peak GPU power draw: 🟠 4x4090 → 1,800W peak (450W × 4) 🟢 4x5090 → 2,300W peak (575W × 4) 🟡 RTX PRO 6000 → 600W peak 🟣 DGX Spark → 240W peak (whole system) The RTX PRO 6000 is the quiet winner. One card, 96GB, matching a 4x5090 rig at roughly a quarter of the power and zero multi-GPU headaches. Best tokens-per-watt by a wide margin. DGX Spark is slow on generation but pulls the least power of any rig here, around 240W for the whole system. Prefill-friendly, memory-rich, wall-socket-friendly. And yes, plenty of people cap their cards. Even then, 4x 4090 or 4x 5090 still pulls well over 1,200W from the GPUs alone.

stevibe

193,239 次观看 • 5 个月前

Gemma4 12B with Unsloth's Quant on DGX Spark Quants: - UD_Q4_K_XL - UD_Q5_K_XL - UD_Q6_K_XL - UD_Q8_K_XL Summary: - Q4: 25.21 tok/s, TTFT 168ms - Q5: 21.7 tok/s, TTFT 182ms - Q6: 17.68 tok/s, TTFT 193.95ms - Q8: 15.22 tok/s, TTFT 221ms

Gemma4 12B with Unsloth's Quant on DGX Spark Quants: - UD_Q4_K_XL - UD_Q5_K_XL - UD_Q6_K_XL - UD_Q8_K_XL Summary: - Q4: 25.21 tok/s, TTFT 168ms - Q5: 21.7 tok/s, TTFT 182ms - Q6: 17.68 tok/s, TTFT 193.95ms - Q8: 15.22 tok/s, TTFT 221ms

stevibe

18,639 次观看 • 4 个月前

the tiebreaker is done. qwen 3.5 27B dense. single RTX 3090. one prompt. zero steering. zero human edits. 1,827 lines across 10 files. 13 minutes. full thinking mode. runs on first load. hermes 4.3 got the same prompt with 2x 3090s and 5x the context it needed. wrote 1,249 lines, left empty files, needed 3 interventions, game was broken on load. same architecture class. same quant. hermes got double the hardware. completely different result. dense wasn't the problem. hermes was. but here's what got me. this model thinks at 27 tok/s. every single token carries 27 billion parameters of reasoning. MoE hit 112 tok/s but only 3B active per token. the dense model is slower and it doesn't matter. watch 13 minutes of autonomous coding on a consumer GPU with zero intervention and tell me speed is what matters. a year ago this wasn't possible. now it runs on hardware you can buy used for $900. no API. no subscription. no cloud. just a 3090 doing what data centers did 18 months ago. full unedited session in the video. every token, every file, every thinking chain. 16 minutes. hit play.

the tiebreaker is done. qwen 3.5 27B dense. single RTX 3090. one prompt. zero steering. zero human edits. 1,827 lines across 10 files. 13 minutes. full thinking mode. runs on first load. hermes 4.3 got the same prompt with 2x 3090s and 5x the context it needed. wrote 1,249 lines, left empty files, needed 3 interventions, game was broken on load. same architecture class. same quant. hermes got double the hardware. completely different result. dense wasn't the problem. hermes was. but here's what got me. this model thinks at 27 tok/s. every single token carries 27 billion parameters of reasoning. MoE hit 112 tok/s but only 3B active per token. the dense model is slower and it doesn't matter. watch 13 minutes of autonomous coding on a consumer GPU with zero intervention and tell me speed is what matters. a year ago this wasn't possible. now it runs on hardware you can buy used for $900. no API. no subscription. no cloud. just a 3090 doing what data centers did 18 months ago. full unedited session in the video. every token, every file, every thinking chain. 16 minutes. hit play.

Sudo su

91,135 次观看 • 7 个月前

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 次观看 • 7 个月前