GPT-OSS 120b

Item: GPT-OSS 120b
Rating: 3
Author: Ali Madjaji

OpenAI · USA

OpenAI's first major open-weights release. Apache 2.0, MoE with 5.1B active over 117B total, MXFP4 quantised to fit a single 80GB GPU. Historic shift for a vendor that built its brand on closed weights.

Sovereignty

Licence: Apache 2.0Commercial: YesTraining data: UndisclosedOrigin: USA

Licence facts

Licence: Apache 2.0
Commercial use: Unrestricted
Derivatives: Allowed
Attribution: Required
Parameters: 117B MoE (5.1B active, MXFP4)
Runtime: Single 80GB GPU (H100 / MI300X)
Training data: Not disclosed
Prompt format: Must use OpenAI 'harmony' format
Last updated: 2025-Q3

Performance & pricing?

Quality index33/100Speed212 tok/sBlended price$0.26/M

Known risks

Non-standard harmony prompt format locks integration work
Training data undisclosed
Reasoning model — chain-of-thought may leak sensitive logic

Sources

→ Model card → Announcement → Harmony format