community pool / model answer

DeepSeek-V4-Flash-0731

chat · deepseek-ai/DeepSeek-V4-Flash-0731

84.8 tok/s · model median · 60 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
CMP 170HX 64GB ×4	8-bit	90.6 tok/s	measured	422	- GB	487,713	vllm
GB10 Grace Blackwell 256GB	8-bit	24.9 tok/s	measured	185	- GB	2,048	vllm
M3 128GB	2-bit	23.3 tok/s	reported	21	87.1 GB	393,216	dwarfstar
Radeon AI Pro R9700 32GB	2-bit	59.4 tok/s	measured	2,004	25.3 GB	131,072	lucebox
Radeon AI Pro R9700 32GB	4-bit	50.5 tok/s	reported	6,063	28.1 GB	32,768	llama.cpp
Radeon AI Pro R9700 32GB ×3	1-bit	37.3 tok/s	measured	1,880	- GB	32,768	llama.cpp
RTX 3090 24GB ×8	8-bit	93.3 tok/s	reported	191	184.4 GB	327,680	vllm
RTX PRO 6000 Blackwell 96GB ×2	8-bit	257.4 tok/s	measured	60	94.5 GB	262,144	vllm
RTX PRO 6000 Blackwell 96GB ×4	4-bit	231.5 tok/s	reported	50	- GB	2,048	vllm
RTX PRO 6000 Blackwell 96GB ×4	8-bit	296.7 tok/s	measured	90	- GB	2,048	vllm
Ryzen AI Max 395 128GB	1-bit	21.8 tok/s	reported	3,305	- GB	782	llama.cpp
Ryzen AI Max 395 128GB	2-bit	28.3 tok/s	reported	15,107	- GB	782	lucebox

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.