GPT-OSS Free

GPT-OSS is OpenAI’s family of open-weight reasoning models, released free under Apache 2.0. The 120B version runs on a single 80GB GPU, the 20B on 16GB, and both can be downloaded from Hugging Face without an API key.

What is GPT-OSS?

In August 2025 OpenAI shipped two dense, mixture-of-experts models branded gpt-oss-120b and gpt-oss-20b. Unlike GPT-5, these are downloadable weights: you get the files, you serve them yourself, and you own the inference stack. The 120B uses roughly 5.1B active parameters per token, which is why OpenAI claims it fits an H100 and still does multi-step reasoning.

The license is Apache 2.0, so commercial use is allowed with no revenue cap. That is the part most teams care about. Fine-tuning is permitted too, though OpenAI did not release a training recipe.

What are the main features of GPT-OSS?

• Two sizes: gpt-oss-120b (117B total, ~5.1B active) and gpt-oss-20b (21B total, ~3.6B active)
• Configurable reasoning effort — low, medium and high — set at request time, not baked into weights
• Native tool use and function calling in the Harmony chat format
• Long context up to 128K tokens
• Runs through Ollama, vLLM, llama.cpp, LM Studio and Transformers
• Apache 2.0 — no per-token fees, no usage dashboard

What makes GPT-OSS worth trying in 2026?

The honest pitch is control. You keep prompts and outputs on your own hardware, which matters for regulated work and for anyone who has been burned by API deprecations. The 20B model runs on a laptop with 16GB VRAM at usable speed, and it is a decent enough assistant for internal tools.

• Zero inference cost per call once you own the GPU — the 120B model served on one H100 handles most chat workloads.
• Reasoning effort is a knob, so you can trade latency for accuracy without swapping models.
• OpenAI’s own safety work is included: the models are trained not to help with bio or chem weapon uplift.

Where does GPT-OSS actually help?

1. Private code review inside a company network where no source can leave the building.
2. A local customer-support copilot that answers from a company wiki.
3. Batch classification of thousands of tickets overnight on a single workstation.
4. Research where you need reproducible outputs and a frozen model version.
5. Teaching students how transformer inference works by watching the model run offline.

Is GPT-OSS free, and who should skip it?

The weights cost nothing, but the hardware does not. Running the 120B model comfortably wants an 80GB accelerator or two 48GB cards; the 20B fits a 4090. If you only have cloud credits and no ops appetite, GPT-5 via API is still simpler. And GPT-OSS trails closed frontier models on hard math and long-horizon agent tasks — OpenAI says so in its own model card.

Quick Summary

• GPT-OSS is OpenAI’s open-weight model family, free under Apache 2.0, shipped in 20B and 120B sizes.
• The 120B model needs about 80GB of GPU memory; the 20B runs on 16GB of VRAM.
• It supports configurable reasoning effort, tool calling and a 128K context window.
• Best for teams that need on-prem inference or zero per-token costs.
• Not the right pick if you need frontier-level math or the lowest possible ops effort.