Key takeaways

• GPT OSS 120B is an open-weight, 117B-parameter Mixture-of-Experts (MoE) language model from OpenAI, designed for high-reasoning, agentic, and general-purpose text-to-text production use cases.
• The model activates 5.1B parameters per forward pass, is optimized for a single H100 GPU with native MXFP4 quantization, and supports configurable reasoning depth, full chain-of-thought access, and native tool use, including function calling, browsing, and structured output generation.
• The model demonstrates strong mathematical reasoning with 96.2% accuracy on MATH-500, particularly excelling in Algebra and Number Theory with detailed explanations, but frequently fails due to strict output formatting requirements.
• A key limitation is its inability to precisely match output formats in benchmarks, causing scoring failures despite mathematically correct underlying logic (e.g., incorrect boxing or term ordering); it also exhibits significant limitations in information extraction, adherence to complex instructions, and a high propensity for hallucination, achieving 0% accuracy on the Sales dataset.
• The model excels in mathematical problem-solving across various domains, achieving high accuracy in Algebra, Number Theory, and Prealgebra tasks, but struggles with strict output formatting and information extraction, leading to a 0% accuracy on the Sales dataset due to hallucination and failure to follow instructions.

Technical specifications

Context length

131k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~175B

Open weights

Yes

Licence

Open

Regions

USA

Vendor

OpenAI

Provider model key

openai/gpt-oss-120b