GPT OSS 120B (flexai)

Artificial Analysis gives gpt-oss-120b an Intelligence Index score of 58, below DeepSeek R1 and Qwen3 235B but ahead of several proprietary models. Its 117B total and 5.1B active parameters provide a smaller deployment footprint than those competing open-weight models. The independent analysis reports a native MXFP4 footprint of 60.8GB, allowing gpt-oss-120b to run on one H100. Its comparatively low active-parameter count improves inference efficiency, while independent evaluations show no single clear area of weakness across the tested categories.

OpenAI released gpt-oss-120b and gpt-oss-20b on August 5, 2025 under an Apache 2.0 license, marking its first open-weight models since GPT-2. The 120b variant carries roughly 117 billion parameters with 5.1 billion active and runs on a single 80GB H100 GPU, while the smaller sibling fits in 16GB. Both can be fine-tuned and deployed locally or hosted on Hugging Face, AWS, Azure, and Databricks. OpenAI's shared benchmarks claim gpt-oss-120b outperforms o4-mini on MMLU, Codeforces, and HealthBench, and the models support tool use, configurable reasoning effort, and chain-of-thought prompting for agentic systems. MLX-optimized builds are available for Apple Silicon Macs, so both variants can run on consumer hardware as a foundation for local desktop LLM applications.

OpenAI's official model card, published on arXiv (2508.10925v1, August 8, 2025), introduces gpt-oss-120b and gpt-oss-20b as open-weight reasoning models released under the Apache 2.0 license. Both are text-only and compatible with the Responses API, intended for agentic workflows with strong instruction following, tool use such as web search and Python code execution, and adjustable reasoning effort. The models are customizable and provide full chain-of-thought with Structured Outputs support. The card details a distinct safety posture for open weights: once released, determined attackers could fine-tune the models to bypass refusals, so developers may need extra safeguards to replicate API-level protections. On gpt-oss-120b, scalable capability evaluations confirmed the default model does not reach OpenAI's indicative High thresholds in the Biological and Chemical, Cyber, and AI Self-Improvement categories of the Preparedness Framework, and adversarial fine-tuning was also investigated.

OpenAI published the model card for gpt-oss-120b and gpt-oss-20b on August 5, 2025, presenting them as open-weight reasoning models released under the Apache 2.0 license and the gpt-oss usage policy. Both are text-only models built for agentic workflows, and the release marks OpenAI's first open-weight models since GPT-2. The document is a model card rather than a system card, since these weights will be embedded in systems maintained by many outside stakeholders. Developers get explicit capabilities in the card: compatibility with the Responses API, tool use such as web search and Python code execution, adjustable reasoning effort, full chain-of-thought, and Structured Outputs support. On safety, OpenAI states that once released, determined attackers could fine-tune these weights to bypass refusals, so developers and enterprises may need extra safeguards beyond API-served protections.

The gpt-oss-120b is a 36-layer mixture-of-experts transformer with 116.83B total parameters, 5.13B active parameters, and a 131K-token context window. It alternates sliding and full grouped-query attention, uses YaRN position scaling, and selects the top four of 128 experts. More than 90% of its parameters are natively quantized to MXFP4 at 4.25 bits per parameter, allowing the model to fit on one 80 GB GPU such as an H100 or MI300X. OpenAI reports strong reasoning, coding, mathematics, tool-calling, and HealthBench results, with especially efficient activation at inference.

OpenAI released gpt-oss-120b on August 5, 2025 under the Apache 2.0 license. The 120B model has 117B total and 5.1B active parameters and can run on a single 80 GB GPU, giving developers an open-weight alternative that can be customized and deployed without API lock-in. OpenAI says gpt-oss-120b reaches near-parity with o4-mini on core reasoning benchmarks and performs strongly in tool use, function calling, chain-of-thought reasoning, and HealthBench. It supports the Responses API, web search, Python execution, structured outputs, and adjustable reasoning effort for agentic applications.

OpenAI’s API documentation identifies gpt-oss-120b as its most powerful open-weight model, with 117B total parameters and 5.1B active parameters that fit on one H100 GPU. The exact model supports a 131,072-token context window and equally large maximum output, while accepting text and producing text. Developers can set reasoning effort to low, medium, or high and inspect the model’s full chain-of-thought. The model supports tools, function calling, web browsing, Python execution, and structured outputs, and is exposed through Chat Completions, Responses, Realtime, and live-session endpoints.