Support · OpenAI gpt-oss
Will it last?
Ownership levelSubstantialnone·limited·partial·substantial·fullAnalytical input B ยท 78.8/100
This page is a projection of the one entry record, the Transparency factor that Support covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
Which domain expands which factor
- AssessUse & modify + Transparency
- ImplementData control + Doesn't fail you
- UseDoesn't fail you
- SupportTransparency
Common problems & fixes
Aggregated common pitfalls (not an exhaustive catalogue):
- Malformed or empty output / leaked reasoning. You are almost certainly not applying the
harmony format - build prompts with the
openai/harmonyrenderer, and never display the raw reasoning channel to users. - Model won't load / dtype errors. MXFP4 needs a recent runtime; upgrade
transformers/vLLM/llama.cpp and load with
torch_dtype="auto"so the library picks MXFP4. - GGUF loads in one runtime but not another. Quant type codes have differed between llama.cpp and Ollama for gpt-oss; use a GGUF built for your specific runtime and re-verify.
- Out-of-memory. Serve the 20b, keep native MXFP4, or reduce max context to shrink the KV cache; the 120b needs an 80GB GPU.
- Non-reproducible results / silent updates. You floated on
main; pin an exact Hugging Face revision and verify checksums.
Versions, changelog & cadence
gpt-oss launched in August 2025 as a single release of two sizes (gpt-oss-120b and
gpt-oss-20b) rather than a rolling semantic-version stream. Individual versions are tracked as
immutable Hugging Face revisions on the openai org, so the revision hash is your
changelog anchor: pin it, and diff against a newer revision if OpenAI republishes. The model
card / arXiv report documents what shipped. (OpenAI's separate gpt-oss-safeguard classifier
family is a different release, not a new gpt-oss version.)
Community & support channels
- GitHub -
openai/gpt-oss(reference implementations, serving/fine-tuning guidance) andopenai/harmony(the response-format renderer) for code-level issues (ev-github). - OpenAI Cookbook guides for gpt-oss deployment and harmony usage.
- Hugging Face model discussion tabs on the
openaiorg (ev-hf-openai) for usage questions.
There is no paid support tier for the open weights - this is community and maintainer support around an open release.
Tracked known issues
Drawn from model-card caveats rather than a formal issue tracker (hence partial):
- Raw chain-of-thought is unsafe to display - it is deliberately unaligned and can contain incorrect or harmful content.
- Safety tuning is removable on open weights - the exact concern behind OpenAI's malicious-fine-tuning study; re-evaluate any derivative.
- Standard hallucination profile for its class - strong reasoning does not eliminate confabulation.
- English-centric - other-language quality is unverified.
How this scores
The ownership factor this domain covers, drawn from the one entry record.
2
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateWeights are open and inspectable and OpenAI published a safety report whose worst-case malicious-fine-tuning methodology was independently reviewed by METR, but the training data and pre-training code are closed, so you adapt rather than reproduce and cannot see how it was made.
How this scores (AOI sub-dimensions)
Provenance4/5how well we can trace and verify what went into the modelVerified openai org on Hugging Face, safetensors-only distribution with checksums, a clear canonical source and no malicious-checkpoint incident on the canonical org (checklist ~6/8).
Governance4/5how accountable and well-documented the publisher isReputable, legally accountable US publisher with a detailed model card / arXiv report, a public reference repository, a documented Preparedness-Framework process and coordinated external safety testing.
What this means for adoptionYou substantially own gpt-oss: ungated Apache-2.0 weights you run, modify and self-host, governed by a one-sentence usage policy (obey applicable law) that is about as light as an open-weight AUP gets. The safety posture is a genuine differentiator - a published worst-case malicious-fine-tuning study whose methodology METR independently reviewed, with OpenAI at least partially addressing each of METR's six high-urgency items. What holds it below full is transparency: the training data and pre-training code are closed, so you adapt rather than reproduce. Adopt it for general open-weight deployment, building to the harmony response format it requires and adding standard input/output guardrails for customer-facing use.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.
OpenAI introduces gpt-oss-120b and gpt-oss-20b as open-weight reasoning models under Apache-2.0 with configurable reasoning effort, native agentic tool use and structured outputs, running on 80GB / 16GB hardware.
gpt-oss-120b & gpt-oss-20b Model Card (arXiv 2508.10925), read verbatim: two open-weight reasoning models on an efficient mixture-of-expert transformer architecture, trained with large-scale distillation and reinforcement learning, releasing all model weights, inference code, tools and tokenizers under an Apache-2.0 license; documents the MoE architecture (~117B/5.1B and ~21B/3.6B, 128/32 experts), 128K context, MXFP4, harmony format and the Preparedness-Framework safety and malicious-fine-tuning evaluations.
gpt-oss weights are hosted on the verified openai organisation on Hugging Face as safetensors under Apache-2.0 with a 128K context window.
gpt-oss LICENSE, read verbatim: a standard Apache License 2.0 (January 2004) boilerplate with the copyright line left as a template placeholder and no bespoke restrictions in the file itself - unconditional commercial use; any use guidance lives separately in the one-sentence USAGE_POLICY, not the licence.
gpt-oss USAGE_POLICY, read verbatim and confirmed complete: the entire policy is a single sentence - "We aim for our tools to be used safely, responsibly, and democratically, while maximising your control over how you use them.
Hugging Face's launch guide documents day-0 gpt-oss support across transformers, vLLM, Ollama and llama.cpp, the native MXFP4 quantization, and the harmony format.
gpt-oss is trained for OpenAI's harmony response format (role hierarchy and separate reasoning/tool/final channels), rendered via the openai/harmony library.
OpenAI gpt-oss safety report PDF, read verbatim: adversarial malicious-fine-tuning on bio/cyber data to simulate a worst case (with external reviewers METR, SecureBio, Daniel Kang); the default model "does not reach High capability" in any of the three Tracked Categories, and OpenAI's Safety Advisory Group concluded that even robust fine-tuning "did not reach High capability in Biological and Chemical Risk or Cyber risk."
METR third-party methodology review of OpenAI's gpt-oss malicious-fine-tuning safety work, read verbatim: "OpenAI at least partially addressed each of our 6 high-urgency items," while flagging that no published High-capability criteria exist; conducted alongside SecureBio and Daniel Kang.
gpt-oss ranks strongly among open-weight models on independent public leaderboards.
The openai/gpt-oss reference repository provides implementations, serving guidance and fine-tuning references for the open weights.
gpt-oss is available in the Ollama model library for local single-command serving.