Welcome GPT OSS, the new open-source model family from OpenAI!

Source: HuggingFace Date: 2025-08-05 URL: https://huggingface.co/blog/welcome-openai-gpt-oss

Summary

Model release: OpenAI’s GPT OSS family — 20B (21B total / 3.6B active, 16GB consumer GPU) and 120B (117B total / 5.1B active, single H100). MoE architecture with MXFP4 4-bit quantization, 128K context, GPT-4o tokenizer. Apache 2.0. Benchmarks: 20B AIME25 pass@1 at 63.3, IFEval 69.5. Supported in Transformers (v4.55.1+), vLLM, llama.cpp. Built-in reasoning with adjustable effort levels and channel-based reasoning/response separation.

Implications

Thread: open-weights ecosystem health / model release cadence. OpenAI releasing open weights under Apache 2.0 is the single biggest signal in this batch — it represents a fundamental strategy shift after years of closed-weights-only positioning. The 20B model running on 16GB consumer GPUs makes this directly competitive with Llama and Qwen for local deployment. MXFP4 native quantization and llama.cpp support on day one suggests this was coordinated to land with maximum ecosystem reach. The channel-based reasoning separation (analysis vs final) is a unique UX feature. Watch whether GPT OSS pulls community attention away from Llama/Qwen or coexists as a fourth major open-weight family.