GPT OSS 20B

OpenAI released gpt-oss-20b on August 5, 2025 as one of two open-weight reasoning models under the Apache 2.0 license. The 20B variant delivers results comparable to OpenAI o3-mini on common benchmarks while running on edge devices with just 16 GB of memory, making it suited for on-device use and local inference. Both gpt-oss models support full chain-of-thought, Structured Outputs, the Responses API, and adjustable reasoning effort for low-latency targets. Tool use, including web search and Python code execution, is integrated for agentic workflows. The release marks OpenAI's first public open-weight GPT-class models since GPT-2. The launch post highlights that gpt-oss-20b was trained using reinforcement learning and techniques informed by OpenAI's frontier systems including o3, and it performs strongly on the Tau-Bench agentic suite and HealthBench, in some cases outperforming proprietary models like o1 and GPT-4o. Safety was treated as foundational, including adversarial fine-tuning evaluation of gpt-oss-120b under OpenAI's Preparedness Framework, with comparable safety behavior reported for the 20B sibling. Custom fine-tuning and full CoT visibility enable developers to adapt the model for specialized applications without costly infrastructure.

Artificial Analysis published an independent benchmark of OpenAI's gpt-oss models on August 6, 2025, assigning gpt-oss-20b an Intelligence Index score of 48 from 20.9B total and 3.6B active parameters. Both variants ship in MXFP4 precision, with the 20B model at just 12.8 GB, light enough for a consumer GPU or 16 GB laptop. The smaller sibling is described as the most intelligent model runnable on a consumer GPU, beating o3-mini in efficiency while trailing larger competitors like DeepSeek R1 0528 and Qwen3 235B. The analysis notes the MoE router selects the top 4 experts per token, with the 20B using 24 layers and 64 query heads under Grouped Query Attention. Both gpt-oss variants are text-only, matching competing open releases from DeepSeek and Alibaba. Inference efficiency from sparse activation enables dozens of output tokens per second on recent MacBooks, supporting local developer use cases without specialized accelerators.

The Hugging Face repository for openai/gpt-oss-20b specifies 21B total parameters with 3.6B active, positioning the variant for lower-latency local and specialized use cases under the permissive Apache 2.0 license. The repo requires the harmony response format and notes MXFP4 quantization of the MoE weights, enabling gpt-oss-20b to run within 16GB of memory. Documented features include configurable reasoning effort, full chain-of-thought, parameter fine-tuning, and agentic capabilities such as function calling, web browsing, Python execution, and Structured Outputs. The repo provides Transformers chat-template usage that auto-applies harmony, plus model.generate instructions for manual harmony formatting or the openai-harmony package, and supports Transformers Serve for an OpenAI-compatible webserver.

OpenAI released its official model card for gpt-oss-120b and gpt-oss-20b on August 5, 2025, framing the pair as open-weight reasoning models distributed under the Apache 2.0 license alongside a dedicated gpt-oss usage policy. The document positions the models as text-only, trained on the harmony response format, and compatible with OpenAI's Responses API for agentic workflows. Key capability highlights include configurable reasoning effort (low, medium, high), full chain-of-thought visibility, native tool use for web search and Python code execution, and Structured Outputs support. OpenAI also published scalable capability evaluations on gpt-oss-120b, confirming the default checkpoint did not cross High capability thresholds in Biological and Chemical, Cyber, or AI Self-Improvement categories under its Preparedness Framework.

Fireworks AI published a technical deep dive on August 5, 2025 covering both gpt-oss-20b and gpt-oss-120b as OpenAI's first open-source LLMs since GPT-2. It positions the models as competitive reasoning systems with adjustable low, mid, and high reasoning levels and strong tool-calling capabilities. Built-in tools like code interpreter and browser, plus user-defined tools, enable multi-turn agentic trajectories for production workflows. The Mixture-of-Experts transformer is described as standard architecturally, with gains driven by training data and reinforcement learning tuning. On benchmarks, the analysis reports gpt-oss-120b surpassing o3-mini and approaching o4-mini, while gpt-oss-20b remains competitive at roughly one-sixth the size. Both models support long context windows and full chain-of-thought reasoning, useful for problem solving and structured agentic workflows.

OpenAI released gpt-oss-20b on August 5, 2025 as an open-weight model with 21 billion total parameters, of which 3.6 billion activate per token, built around a Mixture-of-Experts Transformer architecture that combines dense attention with a local banded sparse attention mechanism and supports a 128,000-token context le According to the model's technical report, its performance is comparable to o3-mini, with strong results on benchmarks such as AIME and HealthBench, and in local deployment tests a quantized GPT-OSS 20B weighing 11.95 GB loaded in 8.3 seconds. The encyclopedia entry also documents distribution milestones: on August 6,