cerebras
Cerebras — Wafer-Scale LLM Inference
Overview
Cerebras Inference serves open-weight models from its own hardware and exposes them at https://api.cerebras.ai/v1, compatible with the OpenAI chat-completions API. It is chosen for speed: time to first token and output rate are far above typical GPU hosting, which matters for interactive chat, autocomplete and multi-step agents.
The model catalog changes quickly and old IDs are removed. At the time of checking (October 2026) the Shared Inference models were:
| Model ID | Notes |
|---|---|
gpt-oss-120b |
about 3,000 tokens/s; context 65k on the free tier, 131k paid; reasoning efforts low/medium/high (default medium) |
qwen-3.8-27b |
about 1,850 tokens/s; context 64k free, 128k paid; reasoning effort none/low/medium/high (default high) |
Other models (for example gemma-4-31b) are Dedicated Inference only. Retired IDs include llama3.1-8b, llama-3.3-70b (the old skill text used llama3.3-70b), qwen-3-32b, qwen-3-235b-a22b-instruct-2507, zai-glm-4.7 and the Llama 4 models; the docs recommend gpt-oss-120b as the replacement for most of them. Always check https://inference-docs.cerebras.ai/models/overview and the deprecations page before hard-coding an ID, or list models at runtime with client.models.list().