models · openai

gpt-oss-20b

gpt-oss-20b — catalogued from third-party evaluation data; no weights held by Sovereign Frontier.

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	openai
Origin	United States
Licence	Apache-2.0
Parameters	20,914,757,184 (20.9B)
Context window	131,072 tokens
Modality	text->text
Architecture	gpt_oss
Layers	24
Hidden size	2880
Attention	GQA (64 heads, 8 KV)
Native precision	U8
Pulls recorded here	10
Downloads reported upstream	6,627,327

Reported scores

MMLU-Pro	74.8
GPQA Diamond	68.8
AIME 2025	89.3
LiveCodeBench	77.7
IFBench	65.1
Humanity's Last Exam	11

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
GeForce RTX 5080	llama.cpp	Q8_0	222 tok/s @ 4,096	candidate	localmaxxing
GeForce RTX 3060	ollama	Q4_K_M	74.0 tok/s @ 2,048	candidate	localmaxxing
Apple M1 Max 64GB	ollama	MXFP4	62.4 tok/s @ 2,048	candidate	localmaxxing
Intel Arc Pro B60	llama.cpp	Q8_0	48.3 tok/s @ 4,096	candidate	localmaxxing
GeForce RTX 3080	ollama	MXFP4	21.0 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
GeForce RTX 3090	llama.cpp	Q4_K_M	180 tok/s @ 512	candidate	localmaxxing
Apple M1 Max 64GB	llama.cpp	Q8_0	83.6 tok/s @ 640	candidate	localmaxxing

From local-ai-registry (MIT), commit 124959d: 11 runs across 10 machines, 0 validated — model revision and runtime pinned, launch accepted — and 11 candidate, their label for useful evidence without a reproducible-launch promise. 1 report figures that contradict each other — prefill below decode, which parallel prefill essentially never is — and are left to the source rather than ranked here.

Measured by local.ai

Hardware	Build measured	Decode @ 8K	Memory @ 8K	Decode mode
NVIDIA GeForce RTX 5090 32GB	—	286 tok/s	30.5 GB	ordinary
NVIDIA GeForce RTX 5090 32GB	mxfp4	286 tok/s	30.5 GB	ordinary
NVIDIA RTX PRO 6000 Blackwell 96GB	—	272 tok/s	90.4 GB	ordinary
NVIDIA RTX PRO 6000 Blackwell 96GB	mxfp4	272 tok/s	90.4 GB	ordinary
NVIDIA GeForce RTX 4090 24GB	—	193 tok/s	23.2 GB	ordinary
NVIDIA GeForce RTX 4090 24GB	mxfp4	193 tok/s	23.2 GB	ordinary
NVIDIA RTX 6000 Ada 48GB	—	181 tok/s	45.8 GB	ordinary
NVIDIA RTX 6000 Ada 48GB	mxfp4	181 tok/s	45.8 GB	ordinary

Measured by local.ai, not by us: 28 runs, 20 more not listed. 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 2026-09-26. 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.