Models / DeepSeek V4 Flash

Best home rig for DeepSeek V4 Flash

Legacy model3 measured resultsfastest plain decode: 95.52 tok/s

Newer model available: DeepSeek V4.1 Flash. DeepSeek V4 Flash is a legacy model; its numbers stay here for reference. source ↗

Real, measured generation speeds for DeepSeek V4 Flash on home hardware, fastest first. Different quants/engines aren't directly comparable.

Build	Quant	Engine	tok/s	Cost	Source
8x RTX 3090 serving DeepSeek V4 Flash (LimeChain vLLM kit)	see source (FP8 KV cache)	vLLM LimeChain SM86 overlay, 1 concurrent request	95.52		github.com ↗
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	IQ3_S	llama.cpp Vulkan	17.4	$7,399 (≈€6,602)	hommie.net ↗
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	IQ3_S, 32K context	llama.cpp Vulkan	12.2	$7,399 (≈€6,602)	hommie.net ↗

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

FAQ

What is the fastest measured hardware for DeepSeek V4 Flash?

8x RTX 3090: 95.52 tok/s generation (vLLM LimeChain SM86 overlay, 1 concurrent request, see source (FP8 KV cache)), from https://github.com/LimeChain/vllm/tree/deepseek-v4-flash-sm86/release/limechain-sm86.

Which GPUs have measured results for DeepSeek V4 Flash?

8x RTX 3090.

Can it run on… 8x RTX 3090