models · deepseek-ai

DeepSeek-V4-Flash-0731

text-generation model harvested from Hugging Face (trending). Catalogued at Quarantine — unverified, pending attestation.

Compare Verify

What this registry has checked

Quarantine. We hold the publisher’s own published hash for these weights and have recorded it against this entry. We have not downloaded and hashed the weights ourselves, so the hash is the publisher’s claim, not our measurement.

Method	hf-published-sha256

Re-check this yourself → Source → Full record →

What is recorded

Publisher	deepseek-ai
Origin	Recorded as a covered nation under 10 U.S.C. § 4872(f) — the PRC, Russia, Iran or North Korea. The registry did not record which.
Licence	mit
Parameters	304,180,418,494 (304.2B)
Context window	1,310,720 tokens
Modality	text->text
Architecture	deepseek_v4
Layers	43
Hidden size	4096
Attention	GQA (64 heads, 1 KV)
Native precision	I8
Pulls recorded here	15,372
Downloads reported upstream	4,469,509

Reported scores

GPQA Diamond	90.8
Humanity's Last Exam	38.6

Reported by a third party and recorded here, not re-run by us. Source: artificialanalysis.ai/api/v2.

Measured on real hardware

None of these runs are ours. Each figure is the reporter’s, linked to their own evidence, under their own trust label. The number shown is single-stream decode — one user, concurrency 1 — because aggregate throughput across many users is a different quantity and reads as far faster than anyone will see.

Reported to local-ai-registry

Hardware	Engine	Build measured	Decode, 1 stream	Label	Reported by
Radeon AI PRO R9700	llama.cpp	MXFP4	52.0 tok/s @ 32,768	candidate	localmaxxing
Radeon AI PRO R9700 ×3 devices	llama.cpp	AD-IQ1_M	48.5 tok/s @ 32,768	candidate	localmaxxing
NVIDIA DGX Spark GB10 ×2 devices	vllm	ModelOpt NVFP4	43.7 tok/s @ 8,192	candidate	mialabs
GeForce RTX 3090	llama.cpp	IQ2_XSS	20.7 tok/s @ 2,048	candidate	localmaxxing
GeForce RTX 5090	llama.cpp	UD-IQ1_S	18.0 tok/s @ 2,048	candidate	localmaxxing
GeForce RTX 5090	llama.cpp	IQ1_S	17.5 tok/s @ 2,048	candidate	localmaxxing
GeForce RTX 5090	llama.cpp	Q8_K_XL	10.4 tok/s @ 2,048	candidate	localmaxxing
Measured outside 2K–32K context — decode speed depends heavily on context, so these are not comparable with the rows above
RTX PRO 6000 Blackwell ×2 devices	vllm	safetensors FP8_E4M3	276 tok/s @ 262,144	candidate	localmaxxing
RTX PRO 6000 Blackwell ×2 devices	sglang	safetensors FP8	57.8 tok/s @ 50	candidate	ppickle1989
GeForce RTX 3090 ×7 devices	llama.cpp	UD-IQ4_XS	50.8 tok/s @ 262,144	candidate	localmaxxing

From local-ai-registry (MIT), commit 124959d: 73 runs across 6 machines, 0 validated — model revision and runtime pinned, launch accepted — and 73 candidate, their label for useful evidence without a reproducible-launch promise. 5 more not listed.

Measured by local.ai

Hardware	Build measured	Decode @ 8K	Memory @ 8K	Decode mode
MacBook Pro M5 Max 128GB · 40-core GPU	UD-Q2_K_XL	29.9 tok/s	99.1 GB	ordinary
NVIDIA DGX Spark 128GB	UD-IQ2_M	19.9 tok/s	93.1 GB	ordinary
NVIDIA DGX Spark 128GB	UD-IQ3_XXS	18.8 tok/s	105.5 GB	ordinary
NVIDIA DGX Spark 128GB	UD-Q2_K_XL	17.5 tok/s	98.8 GB	ordinary
Mac Studio M3 Ultra 96GB · 80-core GPU	UD-Q2_K_XL	7.1 tok/s	99.1 GB	ordinary

Measured by local.ai, not by us: 5 runs. Each names its engine pinned by image digest and the exact serve command, in the full record.

Elsewhere in the registry

All models → Browse with filters → What will run on my hardware? →

Registry record as of 2026-09-01. Identity, licence and parameter count are properties of a release and do not move; tier can, and the verify link above re-checks it against the live registry rather than this page.