gpt-oss-120b & gpt-oss-20b Model Card

gpt-oss-120b and gpt-oss-20b are OpenAI's open-weight reasoning models, released under Apache 2.0 with a usage policy. Both are text-only mixture-of-experts Transformers. The 120b model has 36 layers, 128 experts with 4 active per token, 116.8B parameters and 5.1B active; the 20b model has 24 layers, 32 experts, 20.9B parameters and 3.6B active. Attention alternates between a 128-token sliding window and full attention, with 64 query heads, 8 key-value heads, a learned "attention sink" bias, and 131,072-token context via YaRN.

The expert weights, over 90% of parameters, are quantized to MXFP4 at 4.25 bits per parameter during post-training, so the checkpoints are 60.8 GiB and 12.8 GiB: the 120b model fits one 80 GB GPU and the 20b runs with 16 GB of memory.

Post-training uses chain-of-thought RL similar to o3. The harmony chat format orders roles (system over developer over user over assistant over tool) and separates output into channels: analysis for reasoning, commentary for tool calls and final for the answer. A line in the system prompt sets reasoning effort to low, medium or high. The models were trained to browse, run Python in a Jupyter-like sandbox, and call developer-defined functions.

At high effort the 120b scores 80.1 on GPQA Diamond, 90.0 on MMLU and 62.4 on SWE-Bench Verified, close to o4-mini. Most of the card covers safety: refusal and jailbreak results near o4-mini, weaker instruction-hierarchy results, high hallucination on factual recall, and a test in which OpenAI fine-tuned the 120b model adversarially and found it still below its "High" capability thresholds for biological and cyber risk.

This is the model family a customer often asks about when data cannot leave their environment and they want something close to OpenAI's hosted reasoning models. The card gives the facts that decide whether it works.

It fits. 60.8 GiB of weights on an 80 GB GPU leaves roughly 14 GiB for the KV cache; at about 36 KiB per token for 120b, that is roughly 400K tokens across all concurrent requests. The 20b fits a 16 GB machine with room for about 90K tokens of cache.

Effort is a cost dial, and size matters more than effort. On GPQA with tools, 120b at medium effort (73.5) matches 20b at high effort (74.2) while writing about an eighth of the tokens, by our reading of Figure 3. Higher effort did not help every task: the 20b scored lower at high than medium on Tau-Bench Airline and HealthBench Hard.

It does not know facts reliably. Without browsing, 120b got 16.8% of SimpleQA right and 78.2% wrong. Retrieval or browsing is not optional for factual questions.

Guardrails are your job. The card says developers may need extra safeguards to match OpenAI's hosted protections. The models hold system-prompt instructions less well than o4-mini, and their raw chain of thought is unfiltered and should not be shown to users.