Reasoning weights people can run themselves.

The human problem

Some people and organizations need to customize a model or keep inference on infrastructure they control, but frontier-style reasoning has often required a hosted service.

The prior constraint

Running a capable reasoning model locally commonly required expensive hardware or gave users little ability to modify deployment.

AI’s actual role

Two mixture-of-experts language models designed for tool use, structured outputs, and adjustable reasoning effort.

The documented result

OpenAI states that the 120b model can run within 80 GB of memory and the 20b model within 16 GB; the weights are downloadable from Hugging Face. It reports benchmark results in which 120b approaches o4-mini on core reasoning and 20b is similar to o3-mini on common benchmarks.

Why it may matter

Local or self-hosted use can offer people more deployment choice, including privacy-sensitive settings and adaptation to specialized work.

Limitations

These are open-weight releases, not a complete public release of training data or all model-development details.

OpenAI's performance and safety results are source-vendor evaluations, not independent evidence of benefit in a specific setting.

Open weights can be fine-tuned for harmful purposes; OpenAI explicitly says the models are not intended for medical diagnosis or treatment.

A stated 16 GB memory requirement does not guarantee equally accessible hardware, energy use, or easy setup.

Unresolved questions

How well do locally deployed versions perform in real, safety-critical workflows?

What accessibility and operating-cost barriers remain for people without suitable hardware?

How should harms from malicious fine-tuning be measured and mitigated over time?