community pool / model answer

DeepSeek-V4-Flash

chat · deepseek-ai/DeepSeek-V4-Flash

28.5 tok/s · model median · 36 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
GB10 Grace Blackwell 128GB	8-bit	45.7 tok/s	reported	152	209 GB	200,000	vllm
GB10 Grace Blackwell 256GB	4-bit	26.9 tok/s	reported	495	- GB	393,216	vllm
GB10 Grace Blackwell 256GB	8-bit	36 tok/s	reported	1,202	99.2 GB	500,000	vllm
M3 Ultra 512GB	2-bit	33 tok/s	reported	300	- GB	16,384	mlx
M5 Max 128GB	2-bit	19 tok/s	measured	70,150	- GB	128,000	llama.cpp
Radeon 8060S 103.1GB	2-bit	19 tok/s	reported	659	83 GB	4,096	hipfire
Radeon AI Pro R9700 32GB ×3	1-bit	9.2 tok/s	measured	4,403	81.6 GB	8,192	llama.cpp
Radeon AI Pro R9700 32GB ×4	2-bit	53.7 tok/s	reported	5,602	- GB	2,052	hipfire
RTX 3090 24GB ×4	2-bit	35.3 tok/s	reported	1,283	83.9 GB	2,048	llama.cpp
RTX 5090 32GB ×4	2-bit	212 tok/s	reported	55	124 GB	262,144	vllm
RTX PRO 6000 Blackwell 96GB	3-bit	80.8 tok/s	reported	9,240	94.5 GB	262,144	llama.cpp
RTX PRO 6000 Blackwell 96GB ×2	4-bit	160.9 tok/s	reported	1,130	188.8 GB	500,000	vllm
RTX PRO 6000 Blackwell 96GB ×2	8-bit	261.8 tok/s	reported	4,830	184 GB	262,000	vllm
Ryzen AI Max 395 103GB	2-bit	34.9 tok/s	reported	1,159	- GB	4,096	hipfire
Ryzen AI Max 395 128GB	2-bit	15.2 tok/s	measured	-	- GB	65,536	dwarfstar, hipfire

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.