GPT OSS 120B

OpenAI released gpt-oss-120b on August 5, 2025 as one of two open-weight reasoning models (alongside gpt-oss-20b) under the Apache 2.0 license. The 120b variant achieves near-parity with OpenAI o4-mini on core reasoning benchmarks while running efficiently on a single 80 GB GPU, and it can also run on edge devices with 16 GB of memory for on-device use cases. Both models support tool use, full chain-of-thought, and Structured Outputs via the Responses API. The gpt-oss family was trained using reinforcement learning and techniques informed by OpenAI's frontier internal models including o3. The models demonstrate strong tool use, few-shot function calling, and competitive performance on the Tau-Bench agentic evaluation suite and HealthBench, where they even outperform proprietary models like o1 and GPT-4o. Developers can adjust reasoning effort for low-latency outputs and customize the fully open weights for their own agentic workflows.

Independent architectural analysis confirms gpt-oss-120b is an open-weight autoregressive mixture-of-experts transformer released by OpenAI on August 5, 2025 under Apache 2.0. It has 117B/120B total parameters with 5B/5.1B active parameters, a 131K token context, 36 layers, and a 201,088-token vocabulary. The model alternates banded sliding-window and fully dense attention across 18 layers each, uses Grouped Query Attention with attention-sink tokens, RoPE with YaRN context extension, and gated SwiGLU MoE activations. More than 90% of parameters are MoE weights natively quantized to MXFP4 at 4.25 bits, allowing the 120B model to fit on a single 80 GB GPU such as an NVIDIA H100 or AMD MI300X. OpenAI positions gpt-oss-120b at near-parity with o4-mini on reasoning, matches or exceeds o4-mini on Codeforces, MMLU, HLE, and TauBench tool calling, and beats o4-mini on AIME 2024/2025 mathematics and HealthBench. Weights, inference implementations, tool environments, and tokenizers are all released under Apache 2.0.

OpenAI's API documentation for gpt-oss-120b describes it as the most powerful open-weight model, fitting into a single H100 GPU with 117B total parameters and 5.1B active parameters. It is released under the permissive Apache 2.0 license for experimentation, customization, and commercial deployment, and is downloadable from Hugging Face. The model supports a 131,072-token context window and a 131,072 max output token ceiling. Key features include configurable reasoning effort (low, medium, high), full chain-of-thought access for debugging, parameter fine-tuning support, and native agentic capabilities for function calling, web browsing, Python code execution, and structured outputs. The knowledge cutoff is June 1, 2024, and the model is text-input/text-output only, with reasoning token support available across the Chat Completions, Responses, and Realtime endpoints.

OpenAI's official gpt-oss-120b model card, published August 5, 2025, documents both gpt-oss-120b and gpt-oss-20b as open-weight text-only reasoning models under the Apache 2.0 license. The models are compatible with OpenAI's Responses API, support full chain-of-thought, Structured Outputs, and configurable reasoning effort, and are designed for agentic workflows with web search, Python execution, and function calling. Safety evaluations confirmed the default gpt-oss-120b does not reach High capability thresholds in any Preparedness Framework Tracked Category. OpenAI also adversarially fine-tuned gpt-oss-120b to test Biological/Chemical and Cyber risk escalation, and the Safety Advisory Group concluded it did not reach High capability even under robust attacker-style tuning. OpenAI frames the release as a model card rather than a system card because downstream stakeholders will deploy the open weights in varied systems.