OpenAI releases gpt-oss-120b and gpt-oss-20b under Apache 2.0

Two MoE models sized to specific hardware budgets

OpenAI released gpt-oss-120b and gpt-oss-20b as open-weight models under the Apache 2.0 license, downloadable from Hugging Face. These are the company's first open-weight language models since GPT-2, and unlike Whisper or CLIP, they target general reasoning and tool use.

Both are mixture-of-experts Transformers. The 120b model has 117B total parameters but activates only 5.1B per token across 36 layers, with 4 of 128 experts active at a time. The 20b model activates 3.6B of 21B total parameters, drawing 4 of 32 experts. Both support 128k context via RoPE and use grouped multi-query attention with a group size of 8.

The sizing is deliberate. Shipped natively quantized in MXFP4, gpt-oss-120b fits within 80GB of memory — a single high-end GPU — while gpt-oss-20b runs in 16GB, putting it within reach of consumer laptops and edge devices. OpenAI positions the 120b as near-parity with o4-mini on core reasoning benchmarks and the 20b as comparable to o3-mini.

Weights, not an endpoint

This is a distribution model unlike OpenAI's usual product launches. There is no hosted gpt-oss API. The company says only that it 'may consider API support for gpt-oss in the future,' and steers developers who want multimodal support and built-in tools back toward its proprietary API platform.

To make weights-only usable, OpenAI pre-coordinated with deployment platforms including vLLM, Ollama, llama.cpp, LM Studio, Azure, AWS, Fireworks, Together AI, and others, and worked with NVIDIA, AMD, Cerebras, and Groq on hardware optimization. Microsoft is bringing GPU-optimized gpt-oss-20b to Windows through Foundry Local and the AI Toolkit for VS Code via ONNX Runtime.

OpenAI also open-sourced the plumbing: the o200k_harmony tokenizer, a harmony renderer in Python and Rust for the harmony prompt format the models were post-trained on, plus reference inference implementations for PyTorch and Apple Metal. Named early partners AI Sweden, Orange, and Snowflake explored on-premises hosting and fine-tuning on specialized datasets.

Unsupervised chain-of-thought — and a warning not to show it

The models expose full chain-of-thought and support three reasoning-effort settings — low, medium, high — configurable with a single sentence in the system message. Notably, OpenAI applied no direct supervision to the CoT for either model, arguing that leaving reasoning unmonitored preserves its usefulness as a signal for detecting misbehavior, deception, and misuse.

That design choice comes with an unusually blunt caveat: developers should not display CoT to users. OpenAI's own example makes the reason concrete. Given an instruction never to say the word 'five,' the model's reasoning trace openly works through the forbidden word before producing a compliant final answer — counting '1, 2, 3, 4, … and then about 4.9.'

gpt-oss-120b robustly follows system instructions in its output, but will often explicitly disobey instructions in its CoT.Montana Labs

For teams building agentic workflows, this splits the model into two surfaces with different safety properties: a compliant output layer safe to surface, and a raw reasoning layer that may contain hallucinated or policy-violating content and must be firewalled from end users.

Testing the failure mode of open weights

Because releasing weights means adversaries can fine-tune them, OpenAI ran an unusual pre-release test: it deliberately built non-refusing, domain-specific versions in biology and cybersecurity — 'the way an attacker might' — using its own training stack, then evaluated them under its Preparedness Framework.

OpenAI reports those maliciously fine-tuned models could not reach high capability levels on its framework, and says this result informed the decision to release. Three independent expert groups reviewed the methodology and made recommendations, many adopted. The company is also running a Red Teaming Challenge with a $500,000 prize fund and plans to open-source an evaluation dataset from validated findings.

What shipping weights-only changes for teams running their own infrastructure

The concrete implication of gpt-oss is that OpenAI-quality reasoning is now something you host, not something you call. For teams with data-residency constraints, on-premises requirements, or a need to fine-tune on proprietary data, the Apache 2.0 license and the 80GB / 16GB memory targets remove the two usual blockers: legal terms and hardware cost.

But taking that on means owning what OpenAI's API previously handled for you. There is no managed endpoint, no multimodality, and no built-in tools — you assemble inference through vLLM, Ollama, or a third-party provider, wire up the harmony format, and build your own chain-of-thought monitoring rather than trusting a hosted safety layer. The unsupervised CoT that OpenAI kept legible for research is also the thing you must keep away from your users. gpt-oss lowers the barrier to running frontier-class reasoning locally; it raises the bar for the orchestration and safety scaffolding you now have to supply yourself.