GPT-OSS 120B

Single-80GB target

The release is explicitly positioned around fitting production-grade reasoning into one 80 GB class GPU such as an H100 or MI300X, which is the main operational change versus a typical 100B+ open model.

Configurable reasoning effort

OpenAI exposes low, medium, and high reasoning effort settings so latency and reasoning depth can be traded off at inference time instead of using one fixed behavior.

Native agent features

The model is released with native support for function calling, web browsing, Python execution, and structured outputs rather than treating those as wrapper-level add-ons.

Full chain-of-thought access

The release provides access to the model's reasoning trace for debugging and auditability, though OpenAI notes it is not intended for direct end-user display.

Harmony-only format

Both GPT-OSS models were trained on OpenAI's Harmony response format and are expected to be used with that format rather than a generic chat template.

Model geometry

gpt-oss-120b uses 36 layers, 117B total parameters, 5.1B active parameters per token, 128 total experts, 4 active experts per token, and a 128K context window.

Quantized MoE release

The MoE weights were post-trained in MXFP4, which is the packaging decision that makes the 120B checkpoint practical on a single 80 GB GPU.

Training data and tokenizer

OpenAI describes the training mix as mostly English, text-only data with emphasis on STEM, coding, and general knowledge, tokenized with the open-sourced o200k_harmony tokenizer.

Production general-purpose serving

Best fit when you want one open model that can cover broad assistant, coding, and reasoning workloads without moving to multi-GPU serving first.

High-reasoning workloads

Strong match for use cases that benefit from controllable deeper reasoning rather than the fastest possible low-latency answers.

Fine-tuning and customization

OpenAI explicitly positions the model as fine-tunable, which matters if you want to adapt one large reasoning-capable checkpoint to a narrower production task.

Commercial deployment

The Apache 2.0 license makes it unusually straightforward to experiment, customize, and deploy commercially without copyleft friction.