community pool / model answer

Qwen3.6-27B

chat · Qwen/Qwen3.6-27B

43.7 tok/s · model median · 146 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
Arc Pro B60 24GB ×2	8-bit	9.5 tok/s	reported	1,135	- GB	4,096	llama.cpp
Arc Pro B70 32GB	4-bit	43.5 tok/s	measured	1,156	17.5 GB	131,072	llama.cpp, vllm
Arc Pro B70 32GB	5-bit	26.6 tok/s	measured	311	- GB	204,800	llama.cpp
Arc Pro B70 32GB ×2	4-bit	40.5 tok/s	reported	-	- GB	1,024	llama.cpp
Arc Pro B70 32GB ×2	8-bit	25.7 tok/s	reported	-	- GB	512	llama.cpp
Arc Pro B70 32GB ×3	4-bit	44.2 tok/s	measured	-	- GB	1,024	llama.cpp
Arc Pro B70 32GB ×4	4-bit	33.1 tok/s	reported	-	- GB	1,024	llama.cpp
Arc Pro B70 32GB ×4	8-bit	43.7 tok/s	reported	-	- GB	4,096	vllm
Arc Pro B70 32GB ×4	16-bit	54.2 tok/s	reported	79	17 GB	262,144	vllm
AMD Ryzen 7 7840HS	4-bit	3.5 tok/s	reported	-	- GB	2,048	llama.cpp
GB10 Grace Blackwell 128GB	4-bit	33.3 tok/s	measured	632	117 GB	262,144	llama.cpp, vllm
M1 Max 64GB	8-bit	11.3 tok/s	reported	3,973	- GB	787	llama.cpp
M5 Max 128GB	4-bit	19 tok/s	reported	174	- GB	8,192	lmstudio
Multi-GPU 28GB ×2	4-bit	21.9 tok/s	reported	155	- GB	4,096	llama.cpp
Radeon 8060S Graphics 103GB	4-bit	88.3 tok/s	reported	337	18 GB	4,096	hipfire
Radeon 890M 32GB	4-bit	4.9 tok/s	reported	-	- GB	4,096	llama.cpp
Radeon AI Pro R9700 32GB	4-bit	278 tok/s	reported	103	16.6 GB	2,048	hipfire
Radeon AI Pro R9700 32GB ×3	4-bit	213.7 tok/s	reported	63	16 GB	4,096	hipfire
Radeon AI Pro R9700 32GB ×3	8-bit	17.8 tok/s	measured	768	49.4 GB	262,144	llama.cpp
Radeon AI Pro R9700 34GB	4-bit	254.8 tok/s	reported	111	17.5 GB	4,096	hipfire
Radeon AI Pro R9700 57GB	4-bit	3.6 tok/s	reported	7,848	- GB	2,048	llama.cpp
Radeon RX 7900 XTX 24GB	4-bit	28.4 tok/s	measured	490	- GB	262,144	llama.cpp
RTX 3060 12GB ×2	4-bit	17.6 tok/s	reported	-	18.2 GB	128,000	llama.cpp
RTX 3090 24GB	4-bit	26.2 tok/s	measured	1,374	0 GB	131,072	llama.cpp
RTX 3090 24GB ×2	4-bit	46.7 tok/s	reported	125	- GB	2,048	llama.cpp
RTX 3090 24GB ×2	6-bit	22.7 tok/s	reported	2,101	- GB	2,048	llama.cpp
RTX 3090 24GB ×2	8-bit	40.9 tok/s	reported	2,101	- GB	2,048	vllm
RTX 3090 Ti 24GB	4-bit	140.3 tok/s	reported	710	22 GB	1,024	llama.cpp
RTX 4090 24GB	4-bit	44.2 tok/s	measured	335	- GB	131,072	llama.cpp
RTX 5060 Ti 16GB	4-bit	46 tok/s	reported	453	- GB	2,048	llama.cpp
RTX 5090 32GB	3-bit	73.5 tok/s	measured	36	13.5 GB	8,192	llama.cpp
RTX 5090 32GB	4-bit	70.1 tok/s	measured	37	20 GB	243,040	llama.cpp, vllm
RTX 5090 32GB	8-bit	20 tok/s	reported	200	16 GB	2,048	ollama
RTX PRO 6000 Blackwell 96GB	4-bit	105.7 tok/s	measured	1,150	- GB	15,907	llama.cpp
RTX PRO 6000 Blackwell 96GB	8-bit	72.1 tok/s	measured	1,203	- GB	15,907	llama.cpp
RTX PRO 6000 Blackwell 96GB	16-bit	56.4 tok/s	reported	99	- GB	2,048	vllm
RTX PRO 6000 Blackwell 96GB ×2	16-bit	99.6 tok/s	reported	58	- GB	2,048	vllm
RTX PRO 6000 Blackwell 96GB ×4	16-bit	146.1 tok/s	reported	51	- GB	2,048	vllm
RX 7900 XT 20GB	4-bit	60.6 tok/s	reported	85	- GB	2,048	vllm
RX 7900 XTX 24GB	4-bit	118.2 tok/s	measured	126	16.8 GB	16,384	hipfire, llama.cpp
Ryzen AI Max 395 128GB	4-bit	112 tok/s	reported	290	16.8 GB	2,048	hipfire
Ryzen AI Max 395 128GB	8-bit	16.7 tok/s	reported	1,926	- GB	787	llama.cpp
Ryzen AI Max 395 64GB	4-bit	12.4 tok/s	measured	4,572	17.1 GB	10,240	llama.cpp
Ryzen AI Max 395 64GB	8-bit	7.7 tok/s	reported	4,059	27.8 GB	8,192	llama.cpp
Tesla P100 32GB ×2	4-bit	47.4 tok/s	measured	176	- GB	2,048	llama.cpp
Tesla P100-PCIE-16GB 16GB	3-bit	7.5 tok/s	reported	-	- GB	2,048	llama.cpp
Tesla P100-PCIE-16GB 16GB ×2	5-bit	17.9 tok/s	reported	-	- GB	2,048	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.

back to The wall ->capture / correct ->