community pool / model answer

Qwen3.6-27B

chat · unsloth/Qwen3.6-27B

27.5 tok/s · model median · 41 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
Arc Pro B70 32GB ×2	4-bit	42.1 tok/s	reported	-	- GB	512	llama.cpp
Arc Pro B70 32GB ×3	4-bit	49.4 tok/s	measured	-	- GB	2,048	llama.cpp
Arc Pro B70 32GB ×4	4-bit	39.2 tok/s	measured	-	- GB	1,024	llama.cpp
M3 Ultra 512GB	4-bit	26.3 tok/s	measured	316	- GB	8,192	llama.cpp
M3 Ultra 512GB	8-bit	18.5 tok/s	reported	574	- GB	8,192	lmstudio
M4 Max 48GB	4-bit	21.8 tok/s	reported	1	- GB	4,096	lmstudio
RTX 3090 24GB ×2	4-bit	41.2 tok/s	measured	346	24.3 GB	131,712	llama.cpp, vllm
RTX 3090 24GB ×2	8-bit	45.5 tok/s	reported	804	19 GB	8,192	llama.cpp
RTX 3090 Ti 24GB	4-bit	155.9 tok/s	reported	68	18.4 GB	4,096	llama.cpp
RTX 4090 24GB	3-bit	49.4 tok/s	reported	163	- GB	8,192	llama.cpp
Ryzen AI Max 395 128GB	4-bit	12 tok/s	measured	-	24 GB	65,536	hipfire, llama.cpp, ollama
Ryzen AI Max 395 128GB	8-bit	6.4 tok/s	measured	-	39.3 GB	32,768	llama.cpp

Context tested

The table above is the tested context for each exact rig and quantization cell. Empty fields remain empty rather than being filled with an estimate.