OpenAI gpt-oss

gpt-oss: OpenAI’s first open weights since GPT-2

On August 5, 2025, OpenAI did something it hadn’t done since 2019: it published model weights. gpt-oss-120b and gpt-oss-20b shipped under the Apache 2.0 license — no custom agreement, no acceptable-use rider on the weights themselves — with a full announcement and model card. Whatever you think of the company, this release put genuinely capable reasoning models on pleb hardware, permanently. Weights, once published, don’t un-publish.

What’s in the weights

• gpt-oss-120b: 117B total parameters, Mixture-of-Experts with 128 experts, 5.1B active per token
• gpt-oss-20b: 21B total, 32 experts, 3.6B active per token
• Native MXFP4 quantization: OpenAI ships the weights already quantized to a 4.25-bit block format — this is why the memory numbers below are so friendly
• 128K context window, adjustable reasoning effort (low/medium/high), tool use and agentic workflows as first-class features
• Text-only, English-centric, chain-of-thought visible

OpenAI’s published benchmark claim — and it has held up in independent testing — is that gpt-oss-120b lands near o4-mini on core reasoning benchmarks and gpt-oss-20b matches o3-mini territory. Per the model card, that’s a mid-2025 closed-tier reasoning model you can run without asking anyone’s permission.

Can a pleb run it? Yes — that’s the whole point

Because MXFP4 comes from the factory, you skip the usual quantization math:

• gpt-oss-20b runs in 16 GB of memory. A 16 GB card (4060 Ti 16 GB, used 4080), a used RTX 3090 with room to spare, or any 16 GB+ Apple Silicon Mac. This is the most capable model you can run on the hardware most people already own, and the reason it shows up in every local-LLM tier list.
• gpt-oss-120b runs in 80 GB. One workstation-class 80–96 GB GPU — or, more realistically for home, a big unified-memory box: a 128 GB Strix Halo machine, a Mac with 96 GB+, or the NVIDIA DGX Spark, where the 120b is arguably the appliance’s sweet-spot model. Only 5.1B parameters are active per token, so bandwidth-limited unified-memory hardware still produces usable speeds.

Runtime support is everywhere: ollama pull gpt-oss:20b (or :120b), llama.cpp, LM Studio, vLLM. One caveat worth knowing: gpt-oss speaks OpenAI’s harmony chat format. The mainstream runtimes handle it for you, but if you’re wiring up raw prompts yourself, use the template — mis-formatted prompts are the most common “this model is dumb” bug report.

Why it matters for the sovereign stack

The sovereignty case writes itself: the flagship name in cloud AI published Apache-2.0 weights, and every argument for replacing a cloud AI dependency with a local one got easier to make in a boardroom. “Is local AI actually good?” now has the answer “OpenAI’s own open model, running on your desk.” For Canadian shops with data-residency duties, that pairs directly with the local LLM in Canada playbook.

Know what it is and isn’t, though. Apache 2.0 covers the weights; the training data and pipeline remain closed, so this is open-weight, not open-source-everything — same as almost every family in our license × VRAM comparison. It’s text-only, it’s English-first, and the 20b will still lose to bigger open models on knowledge-heavy tasks. What it does better than anything else at its memory footprint is reasoning and tool use.

Credit where due: gpt-oss exists because the open-weight ecosystem — Llama, Mistral, Qwen, DeepSeek — made open releases table stakes. The plebs won that argument, and these weights are the receipt. If you’re setting up your first rig for it, start with the getting-started steps below, and see how the family compares on the full model roster.