Can you run it?
This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Install & run
gpt-oss checkpoints are safetensors on the verified openai org, shipped natively in MXFP4.
The usual paths all work. Always pin an exact revision.
# Hugging Face transformers.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "openai/gpt-oss-20b" # Or openai/gpt-oss-120b.
rev = "<pinned-commit-sha>" # Pin the revision; do not float on main.
tok = AutoTokenizer.from_pretrained(model_id, revision=rev)
model = AutoModelForCausalLM.from_pretrained(model_id, revision=rev, torch_dtype="auto")
# Ollama (single command).
ollama run gpt-oss:20b
# vLLM (OpenAI-compatible server).
vllm serve openai/gpt-oss-120b --revision <sha>
# llama.cpp / LM Studio (MXFP4 or GGUF).
./llama-cli -m gpt-oss-20b-mxfp4.gguf -p "Hello"
Prefer the safetensors from the openai org and verify per-file checksums after download.
Hardware & VRAM requirements
Only a small fraction of parameters are active per token, so gpt-oss is cheaper to serve than its total-parameter counts suggest. gpt-oss-120b (~117B total, 5.1B active) fits on a single 80GB GPU (H100/MI300-class) in native MXFP4. gpt-oss-20b (~21B total, 3.6B active) runs in about 16GB, which is high-end consumer or edge hardware.
Budget extra headroom for the KV cache at long context lengths (both models support 128K).
Serving stacks
The checkpoints have day-0, first-class runtime support with native MXFP4 kernels, which is why Operational Readiness scores 5.
| Stack | Use it for |
|---|---|
| vLLM | High-throughput serving and OpenAI-compatible API |
| llama.cpp | CPU, edge, GGUF and MXFP4 inference |
| Ollama | Local single-command runs |
| transformers | Prototyping and fine-tuning |
| MLX | Apple-silicon inference |
Hosted endpoints are also available on major providers, so serving can move between self-hosted and managed without changing the client contract.
Safe-deployment controls & Deployment Ceiling
The model-specific gate rows from the safe-deployment playbook are below.
- Supply chain: pin the exact
openairevision hash and verify checksums. Prefer the canonical safetensors (no pickle load). - Prompt layer: render prompts with the harmony response format via
openai/harmony, because the models are trained only for it, and never surface raw chain-of-thought to end users. - Model layer: the models are safety-tuned, but the tuning can be fine-tuned away on open weights. Set conservative decoding defaults.
- Input/output layers: add input/output guardrails and your own guard classifier for any
customer-facing path. The base release ships no guard model, and OpenAI's separate
gpt-oss-safeguardclassifier can fill that role. Add prompt-injection defences and treat retrieved and tool content as untrusted. - Compliance: archive OpenAI's training-content summary and copyright posture in the AI-BOM. The deployment depends on OpenAI for the training-content summary because the data is not published.
Deployment Ceiling: T2 (customer-facing, human-reviewed), conditional. With harmony-correct prompts, input/output guardrails, a guard classifier and no raw-CoT exposure, gpt-oss reaches customer-facing, human-reviewed use. Its strong native tool use makes it a capable agentic backbone. T3+ (bounded autonomous) additionally needs a deterministic action-policy engine, per-action human approval for irreversible operations, and external guard tooling beyond what the release provides. It is held at T2 because the base release ships no companion guard model.
Available quantizations
gpt-oss is natively quantized in MXFP4. The mixture-of-experts weights are released at 4-bit, which lets the 120b fit on one 80GB GPU and the 20b in 16GB. This is a first-class publisher quantization. Community GGUF conversions for llama.cpp and Ollama also circulate. Quant type codes have differed between runtimes (for example llama.cpp and Ollama), so verify that a third-party GGUF loads and behaves correctly before production use.
Fine-tuning & adaptation
The Apache-2.0 weights fine-tune with the standard open stack, PEFT/LoRA/QLoRA via
transformers and TRL on the safetensors checkpoints, and OpenAI's reference repo
(openai/gpt-oss) documents fine-tuning.
Two ceilings apply. The pre-training data and code are not released, so the weights can be fine-tuned and continue-trained but not reproduced from scratch. Fine-tuning can also remove safety behaviour, so re-apply guardrails and re-evaluate any derivative. Placing a derivative on the EU market has provider implications (see Assess).
API / OpenAI-compatible integration
Serve gpt-oss behind vLLM's or Ollama's OpenAI-compatible endpoint for drop-in integration with existing OpenAI-client code.
vllm serve openai/gpt-oss-120b --revision <sha> --port 8000
# Then point any OpenAI SDK at http://localhost:8000/v1.
The one model-specific requirement is the harmony response format. Use the openai/harmony
renderer to construct the role hierarchy and channels instead of passing plain role strings.
With that in place, routing, retries and gateways carry over unchanged.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
ReliabilityIs it reliable and good enough for the job?
StrongThe models run dependably across every major runtime and hosted endpoint, and misuse is well controlled: an independently reviewed worst-case safety study backs the safety tuning.
Doesn't extract your dataDoes running it keep your knowledge and data yours?
StrongSelf-hosted, the ungated Apache weights run entirely on the deployer's own infrastructure with no clawback or telemetry, so data stays with the deployer.
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.