gpt-oss-120b is the model OpenCues ships pointed at: the built-in default is Cerebras serving gpt-oss-120b, with Groq serving the same model as the automatic failover. It earned the slot on three grounds: benchmark accuracy, latency that fits inline surfaces, and the unusual property that two independent fast providers serve the same open weights.
The numbers behind the default
On the project's 137-case fluid-blank suite, Cerebras gpt-oss-120b scores 100% at roughly 262ms per case, the best accuracy-latency combination of any provider tested (Groq's copy of the same model scored 99.3% at 686ms, and was the best cost-per-correct). Cerebras serves it on wafer-scale hardware the docs describe as the fastest inference path OpenCues ships with, around 400-500ms warm for a full dispatch, and it is the only tested provider that sustains medium reasoning effort inside the 500ms word-cue latency budget.
Open weights, two providers, real failover
Because the model is open-weight, Cerebras and Groq both serve the same gpt-oss-120b weights behind the same OpenAI-compatible wire shape; only the URL, the key, and the model name differ. OpenCues exploits that directly: with both keys set, a rate limit or server error on one provider fails over to the other automatically. That is a failover between copies of the same model, not a downgrade to a different one. One honest caveat from the benches: the copies are not byte-identical in behaviour on every surface. Fluid-blank lookups transfer cleanly, but on the transform-blank rewrite suite Groq's copy measured more accurate and Cerebras' faster, a reproducible drift worth knowing if you pin one side.
Optimised around, not just chosen
The default is also the model OpenCues tunes hardest for. Cerebras' automatic prompt prefix caching hits a bench-measured 99.5% cache rate on the ~20k-token static system prompts, saving roughly 300-500ms of first-token time per call (warm calls land around 445ms versus 800-1200ms cold). Reasoning effort is managed per model: Cerebras gpt-oss-120b runs at medium by default and drops to low when you flip max-thinking: off.
Do you have to use it?
No. OpenCues is model agnostic across ten providers, and switching is one scalar in OPENCUES.md (or a natural phrase, with fluid config on). The default just encodes the project's own evidence: for inline surfaces where latency is felt on every keystroke, an open-weight model served fast by two interchangeable providers is a pragmatic centre of gravity. The model selection guide covers when to pick something else.
Related: Cerebras vs Groq · Best model for inline AI · How good is OpenCues? · GitHub repo