OpenAI drops GPT-OSS: a 120B open-source MoE model that fits on a single 80GB GPU

OpenAI just did something it rarely does: it open-sourced a family of models. Meet GPT-OSS, a pair of reasoning-focused models with 21B and 117B total parameters that use a Mixture of Experts (MoE) architecture to keep active parameters low — just 3.6B and 5.1B, respectively. That efficiency trick means the larger 120B variant can squeeze onto a single 80GB GPU, and the 20B version fits into a 16GB consumer card like an RTX 4090. The license is Apache 2.0, with a thin usage policy tacked on, which is about as permissive as you can get from a company that’s historically kept its most powerful models behind an API.

The technical details are genuinely interesting. The models use a 4-bit quantization scheme called mxfp4, but only on the MoE layers, which is how they achieve that dramatic memory reduction. Under the hood, you’ll find SwiGLU activations, RoPE with a 128K context window, and alternating attention layers — some with full context, others with a sliding 128-token window. There’s also a clever learned attention sink per head. They share the same tokenizer as GPT-4o, with a few new tokens to enable compatibility with OpenAI’s Responses API. Inference is supported out of the gate on transformers, vLLM, llama.cpp, and ollama, so the community can get to work immediately.

Hugging Face is rolling out the red carpet, integrating GPT-OSS into its Inference Providers service so you can hit the model through Cerebras or Fireworks AI with the same code you’d use for any OpenAI-compatible endpoint. If you want to run it locally, you’ll need the latest transformers release (v4.55.1), plus accelerate, a kernels library, and Triton 3.4 or better. That Triton version is key — it unlocks mxfp4 support on older CUDA architectures like Ampere and Ada, which covers most of the GPUs developers actually own. Without those libraries, the model falls back to bfloat16 and suddenly the 20B model wants 48GB of VRAM instead of 16GB.

OpenAI frames this as a meaningful step toward its mission of making AI broadly accessible, and for once, the release backs up the rhetoric. The combination of Apache 2.0 licensing, aggressive quantization that works on hardware people already have, and immediate ecosystem support makes this feel less like a marketing exercise and more like a genuine olive branch to the open-source community. Whether these models actually perform at a level that justifies the attention is the next question — but the plumbing is in place for developers to find out fast.