OpenAI Ships Open-Weight Reasoning Models, First Since GPT-2

OpenAI just did something it hasn’t done since 2019: release an open-weight language model. The gpt-oss family lands with two sizes—a 120-billion-parameter beast and a lean 20-billion-parameter sibling—both under the Apache 2.0 license. If you’ve been watching the open-source community chase proprietary performance, this is a direct shot across the bow. The larger model, gpt-oss-120b, nearly matches the company’s own o4-mini on reasoning benchmarks while fitting on a single 80 GB GPU. That’s not just competitive; it’s a practical signal that frontier reasoning doesn’t require a server farm anymore.

The real eyebrow-raiser is the 20b variant. It trades blows with o3-mini on common benchmarks but sips just 16 GB of memory—low enough to run on edge devices and laptops. For developers who’ve been cobbling together tool-use pipelines with smaller open models, having something this capable at that footprint changes the math. Both models lean heavily on chain-of-thought reasoning and function calling, and they didn’t just clear standard evals. On HealthBench, they outpaced proprietary heavyweights like GPT-4o and o1. That’s a medical benchmark, not a toy dataset, which makes the result genuinely interesting.

Under the hood, the pair uses a mixture-of-experts architecture to keep active parameters low—5.1B per token for the 120b model, 3.6B for the 20b. Context windows stretch to 128k tokens, and they’re trained on a text-heavy dataset skewed toward STEM and coding. OpenAI also open-sourced the tokenizer, a superset of what powers o4-mini and GPT-4o. Training mixed reinforcement learning with techniques informed by frontier systems like o3. Safety testing included an adversarial fine-tuning pass under the Preparedness Framework, plus external expert review—a nod to the chaos that often follows open releases.

Early partners like AI Sweden, Orange, and Snowflake are already kicking the tires on real-world deployments, from on-prem hosting to custom fine-tuning. The move gives developers a genuine spectrum: use the API for convenience and scale, or grab these weights and run them on your own iron. For a company that’s been synonymous with closed doors, shipping open models this capable feels less like charity and more like acknowledgment that the ground has shifted. Whether this accelerates the commoditization of frontier reasoning or just gives enterprises a safer on-ramp to self-hosting remains to be seen, but it certainly makes the open-weight landscape a lot more crowded at the top.