Six years after GPT-2, OpenAI finally dropped weights into the open on August 5. Two models: gpt-oss-120b and gpt-oss-20b, both under Apache 2.0, both built for agentic work, both downloadable from Hugging Face today. The question worth asking is not whether this is historically significant - it is - but whether these models earn a slot in a real inference stack right now.
What gpt-oss-120b and gpt-oss-20b actually are
Each model is a transformer using mixture-of-experts to reduce active parameter count. gpt-oss-120b activates 5.1 billion parameters per token from 117 billion total; gpt-oss-20b activates 3.6 billion from 21 billion total. Both use the same chat template and are fine-tuned variants of a shared base. They natively support context lengths up to 128k tokens and were trained on a mostly English, text-only dataset with emphasis on STEM, coding, and general knowledge.
Similar to OpenAI's o-series models, both gpt-oss models let users toggle between three reasoning effort levels - low, medium, and high - all through a single system prompt. That is not a small thing: it means one deployed model handles cheap quick lookups and expensive deep reasoning without swapping weights.
The models are not available through the OpenAI API. Users download them for free from Hugging Face and run them locally - no cloud connection required. They are available under Apache 2.0 and can run through third-party platforms including Azure, Ollama, AWS, Together AI, Vercel, Cloudflare, and Databricks.
Microsoft is also distributing optimized versions of gpt-oss-20b to Windows devices via ONNX Runtime, available through Foundry Local and Visual Studio Code's AI Toolkit.
How gpt-oss benchmarks - and where it surprises
gpt-oss-120b and gpt-oss-20b generally perform on par with o4-mini. Both models perform within 1-2 points of o4-mini on standard disallowed content evaluations and tend to significantly outperform o4-mini on production benchmarks.
On HealthBench, the 120b model at high reasoning performs nearly as well as o3 and substantially better than GPT-4o, o1, o3-mini, and o4-mini. The 20b model performs slightly better than o1, despite being significantly smaller.
Here is where it gets counterintuitive: on AIME 2025 competition math, both models perform surprisingly well - and the 20b model actually edges out the 120b with 98.7% versus 97.9%. An independent comprehensive evaluation found the same inversion across other tasks. GPT-OSS-20B outperforms GPT-OSS-120B despite being smaller, challenging scaling assumptions in sparse MoE architectures. That is not a rounding error; it suggests the 20b's training regime or routing structure happens to fit certain reasoning tasks better at its scale.
For inference speed on hosted providers: gpt-oss-120b generates 176 tokens per second versus 164.8 for the 20b. The 20b costs roughly $0.08 per million tokens compared to $0.20 for the 120b at a 7:2:1 cache hit/input/output ratio. At that delta, the 20b is the obvious first deployment choice for any team running volume.
The gap nobody in the coverage is talking about
The benchmark tables OpenAI leads with are English-first. The comprehensive third-party evaluation that went deeper tells a different story: both gpt-oss variants scored poorly on Chinese-language benchmarks - 28% for the 20b and 20% for the 120b on C-Eval, well below the 45% threshold researchers use as a baseline. Models with language-specific optimization, such as Qwen 3, achieve substantially higher accuracy. This gap indicates that general-purpose pretraining alone is insufficient for robust multilingual capability.
That 20% figure for the 120b is not a minor miss. It means the larger model is worse at Chinese than the smaller one, and both sit far behind the Chinese-lab open-weight alternatives that have spent training compute on non-English corpora. Any team serving multilingual users or building on top of Chinese-language data should treat gpt-oss as English-first infrastructure, not a general-purpose deployment.
What genuinely changed versus the incremental story
The hype framing - "OpenAI finally goes open-source" - is technically accurate but slightly misleading. This is the first time OpenAI has published an open-weight model since GPT-2 in 2019. But the models are competitive with o4-mini and o3-mini, not with GPT-5 or o3. The gpt-oss series is capable and competitive with OpenAI's smaller proprietary models, but it is not state-of-the-art in the same way as GPT-5. It is mostly about giving developers, researchers, and companies the ability to run models locally, fine-tune them for specific domains, and experiment without starting from scratch.
That framing is actually more useful for deciding whether to deploy them. If your team is building an agent that needs reliable function calling, 128k context, and adjustable reasoning depth - and you want to run it on-premises with full data control - these models are a real option. The gpt-oss models have particularly strong performance on coding and tool-use tasks. On Codeforces, SWE-Bench, and τ-bench retail, gpt-oss-120b comes close to o4-mini in performance.
What changed is the deployment model, not the capability ceiling. You can now run an o4-mini-class reasoning model on a single H100 - or close to it - without routing traffic through an API, without per-token billing, and with the ability to fine-tune on proprietary data under a license that does not complicate your IP situation.
A teammate like Beagle - running inside Slack and pulling from internal docs - is a reasonable use case for exactly this class of model: high function-calling reliability, adjustable reasoning depth based on query complexity, and deployable in a private cloud with no data leaving the perimeter.
Beagle in action#engineering, 10:42amOpenAI gpt-oss: common questions
What hardware does gpt-oss-120b require to run?
The gpt-oss-120b model can run on a single 80GB GPU like an NVIDIA H100, while the gpt-oss-20b can operate on systems with as little as 16GB of memory. Both support common inference stacks including vLLM, Ollama, and llama.cpp. For 4-bit quantized variants, lower-memory consumer setups are feasible with some throughput trade-off.
How does gpt-oss compare to DeepSeek and Qwen on coding?
On SWE-Bench and Codeforces, gpt-oss-120b approaches o4-mini-level performance. But independent evaluations found that Mistral Small 4 outperforms gpt-oss-120b on LiveCodeBench while producing 20% less output , and the Chinese-lab alternatives have stronger coding benchmarks in absolute terms at equivalent or lower active parameter counts. gpt-oss is competitive but not the ceiling for open-weight coding tasks.
Can I fine-tune gpt-oss on private data?
The trained weights are publicly available under the permissive Apache 2.0 license and gpt-oss usage policy, which means you can download the models and run them on your own infrastructure. Fine-tuning is supported via standard tooling. The main quirk to learn is the Harmony format required by the chat template, which is well documented in OpenAI's cookbook.
Is gpt-oss safe to deploy as an agent?
Safety is particularly important for open models. OpenAI ran the models through comprehensive safety training and also tested an adversarially fine-tuned version of gpt-oss-120b under their Preparedness Framework. The default models passed those evaluations. The caveat is that once weights are public, adversarial fine-tuning is possible - OpenAI's own safety paper documents this risk and publishes what they found when they tried it themselves.
Why is gpt-oss-20b sometimes better than gpt-oss-120b?
On at least two benchmarks - AIME 2025 math and several general-capability tasks - the 20b outperforms the 120b. This appears to be a consequence of how mixture-of-experts models route tokens at different scales, not a calibration error. For production deployments, it is worth running your own evals at the 20b tier before defaulting to the larger model: you may get better results for less than half the inference cost.