Cerebras Inference Lists Publicly Available Models and Compression Policy

Summary

Cerebras Inference’s model catalog covers models served through its public endpoints, which are available on free-trial and pay-as-you-go tiers subject to rate limits and pricing. The page lists OpenAI GPT OSS with the model ID gpt-oss-120b, 120 billion parameters, 65k context on the free tier and 131k on the paid tier, and an advertised speed of about 3,000 tokens per second. It also lists Qwen 3.8 27B, with the ID qwen-3.8-27b, 27 billion parameters, 64k/128k free-versus-paid context limits, and about 1,500 tokens per second. Cerebras says additional model families, reserved capacity, higher throughput, and production SLAs are available through Dedicated Endpoints. The catalog states that its public models are original, unpruned versions; the company does not currently serve pruned models on shared public endpoints. Cerebras researches pruning methods including REAP, but says those pruned models are shared on Hugging Face for research and experimentation rather than through its production API. For storage, the company uses selective weight-only quantization at partial 16-bit, 8-bit, and 4-bit precision, while sensitive layers remain at full precision and are dequantized during execution. Activations, attention, and the key-value cache remain full precision and unquantized. Cerebras also says it will not change the architecture of existing endpoints without notice. Any future pruned or otherwise differently compressed versions would be exposed as separately named endpoints so users can distinguish them from the original models.