gpt-oss-120b and gpt-oss-20b

2025-08-05 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI releases gpt-oss-120b and gpt-oss-20b, its first open-weights language models since GPT-2, under the Apache 2.0 license, designed for reasoning, tool use and deployment on commodity hardware.

What changes

  • gpt-oss-120b reaches near parity with o4-mini on reasoning benchmarks and fits on a single 80 GB GPU.
  • gpt-oss-20b delivers results close to o3-mini and runs with 16 GB of memory, including on local devices.
  • Adjustable reasoning effort (low, medium, high), full unsupervised chain of thought, function calling, web browsing, Python execution and structured outputs.
  • Downloadable and customizable weights (fine-tuning), deployable on-premises for data privacy.

OpenAI releases gpt-oss-120b and gpt-oss-20b, its first open-weights language models since GPT-2, under the Apache 2.0 license, designed for reasoning, tool use and deployment on commodity hardware. ### What changes - gpt-oss-120b reaches near parity with o4-mini on reasoning benchmarks and fits on a single 80 GB GPU. - gpt-oss-20b delivers results close to o3-mini and runs with 16 GB of memory, including on local devices. - Adjustable reasoning effort (low, medium, high), full unsupervised chain of thought, function calling, web browsing, Python execution and structured outputs. - Downloadable and customizable weights (fine-tuning), deployable on-premises for data privacy. ### Specifications - Mixture-of-experts (MoE) Transformer architecture: 117 billion total parameters and 5.1 billion active per token for the 120b; 21 billion and 3.6 billion for the 20b. - Native context up to 128,000 tokens, text input and output, knowledge cutoff June 2024, MXFP4 quantization. - Availability: Hugging Face, Azure, AWS, Databricks, Ollama, vLLM, LM Studio, Cloudflare, OpenRouter and others; no OpenAI API at launch. - Compatible with the Responses API; new o200k_harmony tokenizer published. ### Pricing Free weights under the Apache 2.0 license, with no OpenAI pricing. The cost depends on the hardware or the host chosen. ### Benchmarks Scores published by the vendor: the post states in qualitative terms that gpt-oss-120b matches or exceeds o4-mini on Codeforces, MMLU, HLE and TauBench, and exceeds it on HealthBench and AIME 2024 and 2025. The values are in charts, not included here. ### Key takeaways These models offer a sovereign, self-hosted option for use cases with sensitive data. OpenAI specifies that the models on its platform remain the best option for multimodal and built-in tools.

Links

Tags : Open Source, LLM, OpenAI, gpt-oss, Raisonnement, Apache 2.0