Cerebras Inference model catalog (shared-inference models, context, vendor speed figures)

Fetched 2026-10-14. Shared Inference (Free Trial / Pay-as-you-go) hosts exactly two models: gpt-oss-120b (120B, 65k ctx free / 131k paid, vendor-stated ~3000 tok/s) and qwen-3.8-27b (27B, 64k/128k, ~1850 tok/s). Larger catalog sits behind Dedicated Inference (reserved capacity, sales agreement). Page also states all shared models are unpruned, with selective weight-only quantization at storage and full-precision activations/KV.