GPT OSS 120B is an open-weight text language model from the gpt-oss family that multiple inference hosts expose as a shared production SKU. According to vendor public tables retrieved 18 August 2026, Cerebras lists the model near 3,000 tokens per second while Groq lists the same weights near 500 tokens per second, illustrating how dramatically throughput varies by hardware even when the underlying parameters are identical. The open-weight nature means the same artifact can be deployed across different serving stacks, with each host setting its own balance of speed and price. This makes the model well suited to developers who want a transparent, self-contained 120-billion-parameter text generator and are willing to choose a host that matches their latency and cost priorities.
In Pipecat's third-party voice-readiness benchmark of 46 model configurations across 7 providers, the Groq-served GPT OSS 120B posted a 98 ms median time to first answer but only an 86.3% strict pass rate on 30-turn scripted conversations covering tool use, instruction following, and knowledge-base grounding. That places the configuration among the fastest to first token in the field, well inside the ~700 ms voice budget, yet its error rates on instructions (13.0%) and tool calls (9.3%) are higher than many slower peers. The model therefore fits best in latency-sensitive text pipelines where fast first response matters more than turn-level accuracy, such as live autocomplete, streaming classification, or thin tool wrappers, rather than in long multi-step agent loops where instruction drift accumulates.
