Each column is one machine. The title and the links under it open that box's report and test pages (C, Python, Bash, research). A model name or score opens that model's tests on that machine. /126 is the full suite; a smaller denominator means the sweep is still incomplete. Tok/s is not comparable across columns. Side-by-side scores: compare. Generated 2026-10-02 10:57.
CPU M5 Max, 16-core CPU
GPU 40-core GPU · Metal
memory 128 GB unified
runtime MLX, llama.cpp Metal
| model | type | coding | tok/s |
|---|---|---|---|
| aya-23 35b llama.cpp Metal | dense | 66/126 | 11 |
| devstral-2 24b llama.cpp Metal | dense | 94/126 | 24 |
| gemma-4-26b llama.cpp Metal | MoE | 102/126 | 61 |
| glm-4.7-flash llama.cpp Metal | MoE | 94/126 | 48 |
| gpt-oss-120b llama.cpp Metal | MoE | 106/126 | 62 |
| gpt-oss-20b llama.cpp Metal | MoE | 104/126 | 83 |
| k2-horizon 36B-A4B llama.cpp Metal | MoE | 86/126 | 47 |
| kimi-k3 (referee) cloud | — | 44/44 | — |
| laguna-s 2.1 117B llama.cpp Metal | MoE | 101/126 | 40 |
| laguna-xs.2 llama.cpp Metal | MoE | 100/126 | 81 |
| mistral-small-4 119B llama.cpp Metal | MoE | 91/126 | 52 |
| nemotron-3-super 120B llama.cpp Metal | MoE | 102/126 | 20 |
| qwen3-coder-30b llama.cpp Metal | MoE | 86/126 | 81 |
| qwen3-coder-next 80B llama.cpp Metal | MoE | 99/126 | 55 |
| qwen3.5-122b-a10b llama.cpp Metal | MoE | 90/126 | 27 |
| qwen3.5-27b llama.cpp Metal | dense | 87/126 | 14 |
| qwen3.5-35b llama.cpp Metal | MoE | 96/126 | 58 |
| qwen3.6-27b llama.cpp Metal | dense | 99/126 | 15 |
| qwen3.6-35b llama.cpp Metal | MoE | 96/126 | 63 |
| qwen3.8-27b llama.cpp Metal | dense | 102/126 | 14 |
| qwen3.8-flash-next 125B llama.cpp Metal | MoE | 108/126 | 33 |
| seed-oss 36b llama.cpp Metal | dense | 79/126 | 11 |
Full report for this machine → Tok/s is only comparable inside this column. 22 models.
CPU Ryzen 9 9950X3D, 16-core
GPU RTX PRO 6000 Blackwell, 96 GB · CUDA sm_120
memory 185 GB RAM + 96 GB VRAM
runtime llama.cpp CUDA, vLLM, SGLang, Ollama
| model | type | coding | tok/s |
|---|---|---|---|
| aya-23 35b llama.cpp CUDA | dense | 61/126 | — |
| devstral-2 24b llama.cpp CUDA | dense | 96/126 | — |
| gemma-4-26b llama.cpp CUDA | MoE | 105/126 | — |
| glm-4.7-flash llama.cpp CUDA | MoE | 94/126 | — |
| gpt-oss-120b llama.cpp CUDA | MoE | 111/126 | — |
| gpt-oss-20b llama.cpp CUDA | MoE | 102/126 | — |
| k2-horizon 36B-A4B llama.cpp CUDA | MoE | 91/126 | — |
| kimi-k3 (referee) cloud | — | 44/44 | — |
| laguna-s 2.1 117B llama.cpp CUDA | MoE | 108/126 | — |
| laguna-xs-2.1 llama.cpp CUDA | MoE | 105/126 | — |
| mistral-small-4 119B llama.cpp CUDA | MoE | 87/126 | — |
| nemotron-3-super 120B llama.cpp CUDA | MoE | 105/126 | — |
| qwen3-coder-30b llama.cpp CUDA | MoE | 93/126 | — |
| qwen3-coder-next 80B llama.cpp CUDA | MoE | 99/126 | — |
| qwen3.5-122b-a10b llama.cpp CUDA | MoE | 93/126 | — |
| qwen3.5-27b llama.cpp CUDA | dense | 93/126 | — |
| qwen3.5-35b llama.cpp CUDA | MoE | 105/126 | — |
| qwen3.6-27b llama.cpp CUDA | dense | 99/126 | — |
| qwen3.6-35b llama.cpp CUDA | MoE | 96/126 | — |
| qwen3.8-27b llama.cpp CUDA | dense | 102/126 | — |
| qwen3.8-flash-next 125B llama.cpp CUDA | MoE | 117/126 | — |
| seed-oss 36b llama.cpp CUDA | dense | 82/126 | — |
Full report for this machine → Tok/s is only comparable inside this column. 22 models.
CPU 20-core Grace (Cortex-X925 / Cortex-A725)
GPU NVIDIA GB10 · CUDA sm_121
memory 128 GB unified
runtime llama.cpp CUDA
| model | type | coding | tok/s |
|---|---|---|---|
| aya-23 35b llama.cpp CUDA | dense | 73/126 | 9 |
| devstral-2 24b llama.cpp CUDA | dense | 98/126 | 14 |
| gemma-4-26b llama.cpp CUDA | MoE | 102/126 | 56 |
| glm-4.7-flash llama.cpp CUDA | MoE | 96/126 | 67 |
| gpt-oss-120b llama.cpp CUDA | MoE | 107/126 | 49 |
| gpt-oss-20b llama.cpp CUDA | MoE | 102/126 | 82 |
| k2-horizon 36B-A4B llama.cpp CUDA | MoE | 88/126 | 48 |
| kimi-k3 (referee) cloud | — | 44/44 | — |
| laguna-s 2.1 117B llama.cpp CUDA | MoE | 108/126 | 28 |
| laguna-xs-2.1 llama.cpp CUDA | MoE | 101/126 | 88 |
| mistral-small-4 119B llama.cpp CUDA | MoE | 83/126 | 36 |
| nemotron-3-super 120B llama.cpp CUDA | MoE | 105/126 | 16 |
| qwen3-coder-30b llama.cpp CUDA | MoE | 90/126 | 68 |
| qwen3-coder-next 80B llama.cpp CUDA | MoE | 108/126 | 52 |
| qwen3.5-122b-a10b llama.cpp CUDA | MoE | 93/126 | 22 |
| qwen3.5-27b llama.cpp CUDA | dense | 96/126 | 12 |
| qwen3.5-35b llama.cpp CUDA | MoE | 96/126 | 67 |
| qwen3.6-27b llama.cpp CUDA | dense | 99/126 | 12 |
| qwen3.6-35b llama.cpp CUDA | MoE | 102/126 | 67 |
| qwen3.8-27b llama.cpp CUDA | dense | 102/126 | 12 |
| qwen3.8-flash-next 125B llama.cpp CUDA | MoE | 111/126 | 28 |
| seed-oss 36b llama.cpp CUDA | dense | 81/126 | 9 |
Full report for this machine → Tok/s is only comparable inside this column. 22 models.
CPU 20-core Grace (Cortex-X925 / Cortex-A725)
GPU NVIDIA GB10 · CUDA sm_121
memory 128 GB unified
runtime llama.cpp CUDA
| model | type | coding | tok/s |
|---|---|---|---|
| aya-23 35b llama.cpp CUDA | dense | 73/126 | 9 |
| devstral-2 24b llama.cpp CUDA | dense | 98/126 | 14 |
| gemma-4-26b llama.cpp CUDA | MoE | 102/126 | 56 |
| glm-4.7-flash llama.cpp CUDA | MoE | 96/126 | 67 |
| gpt-oss-120b llama.cpp CUDA | MoE | 107/126 | 54 |
| gpt-oss-20b llama.cpp CUDA | MoE | 102/126 | 81 |
| k2-horizon 36B-A4B llama.cpp CUDA | MoE | 88/126 | 51 |
| kimi-k3 (referee) cloud | — | 44/44 | — |
| laguna-s 2.1 117B llama.cpp CUDA | MoE | 108/126 | 35 |
| laguna-xs-2.1 llama.cpp CUDA | MoE | 102/126 | 94 |
| mistral-small-4 119B llama.cpp CUDA | MoE | 83/126 | 43 |
| nemotron-3-super 120B llama.cpp CUDA | MoE | 105/126 | 16 |
| qwen3-coder-30b llama.cpp CUDA | MoE | 90/126 | 91 |
| qwen3-coder-next 80B llama.cpp CUDA | MoE | 108/126 | 51 |
| qwen3.5-122b-a10b llama.cpp CUDA | MoE | 93/126 | 22 |
| qwen3.5-27b llama.cpp CUDA | dense | 96/126 | 13 |
| qwen3.5-35b llama.cpp CUDA | MoE | 96/126 | 69 |
| qwen3.6-27b llama.cpp CUDA | dense | 99/126 | 11 |
| qwen3.6-35b llama.cpp CUDA | MoE | 102/126 | 60 |
| qwen3.8-27b llama.cpp CUDA | dense | 102/126 | 11 |
| qwen3.8-flash-next 125B llama.cpp CUDA | MoE | 111/126 | 34 |
| seed-oss 36b llama.cpp CUDA | dense | 81/126 | 9 |
Full report for this machine → Tok/s is only comparable inside this column. 22 models.
CPU Ryzen AI MAX+ 395, 16-core / 32-thread
GPU Radeon 8060S gfx1151, 40 CU · Vulkan/RADV
memory 128 GB LPDDR5 (BIOS: 64 GB VRAM + 62 GB RAM; Vulkan VRAM+GTT ≈ 95 GB)
runtime llama.cpp Vulkan
| model | type | coding | tok/s |
|---|---|---|---|
| aya-23 35b llama.cpp Vulkan | dense | 64/126 | 10 |
| devstral-2 24b llama.cpp Vulkan | dense | 96/126 | 15 |
| gemma-4-26b llama.cpp Vulkan | MoE | 105/126 | 47 |
| glm-4.7-flash llama.cpp Vulkan | MoE | 103/126 | 64 |
| gpt-oss-120b llama.cpp Vulkan | MoE | 109/126 | 51 |
| gpt-oss-20b llama.cpp Vulkan | MoE | 104/126 | 72 |
| k2-horizon 36B-A4B llama.cpp Vulkan | MoE | 91/126 | 48 |
| kimi-k3 (referee) cloud | — | 44/44 | — |
| laguna-s 2.1 117B llama.cpp Vulkan | MoE | 107/126 | 34 |
| laguna-xs-2.1 llama.cpp Vulkan | MoE | 102/126 | 86 |
| mistral-small-4 119B llama.cpp Vulkan | MoE | 82/126 | 42 |
| nemotron-3-super 120B llama.cpp Vulkan | MoE | 104/126 | 17 |
| qwen3-coder-30b llama.cpp Vulkan | MoE | 87/126 | 85 |
| qwen3-coder-next 80B llama.cpp Vulkan | MoE | 105/126 | 56 |
| qwen3.5-122b-a10b llama.cpp Vulkan | MoE | 93/126 | 22 |
| qwen3.5-27b llama.cpp Vulkan | dense | 93/126 | 13 |
| qwen3.5-35b llama.cpp Vulkan | MoE | 95/126 | 59 |
| qwen3.6-27b llama.cpp Vulkan | dense | 100/126 | 12 |
| qwen3.6-35b llama.cpp Vulkan | MoE | 94/126 | 59 |
| qwen3.8-27b llama.cpp Vulkan | dense | 105/126 | 12 |
| qwen3.8-flash-next 125B llama.cpp Vulkan | MoE | 115/126 | 28 |
| seed-oss 36b llama.cpp Vulkan | dense | 86/126 | 10 |
Full report for this machine → Tok/s is only comparable inside this column. 22 models.