Cerebras Model Catalog (Website)

Cerebras provides GPT OSS and Qwen 3.8 models via public endpoints. GPT OSS operates at a speed of approximately 3000 tokens/s, supporting a context window of 64k on the free tier and 128k on the paid tier. Qwen 3.8 also has similar speed and context limits.

Model Compression and Quality Control

Cerebras emphasizes that all models served on public endpoints are in their original (unpruned) state. Since pruning techniques alter model structure and can cause performance degradation, REAP pruned models developed for research are shared only on Hugging Face and are not served via API.

Instead, Cerebras applies selective weight-only quantization at storage time. This method stores weights in FP8 or FP4 to improve memory efficiency, but immediately dequantizes them to high precision during computation to maintain quality. Activations and KV cache are always maintained at full precision.