Qwen 3.8 27B, a new model with 27 billion parameters, is now available on Cerebras, achieving a speed of 1500 tokens per second.

This model operates with a context length of 64k for free users and 128k for paid users. Cerebras maintains high-quality model performance by using selective weight-only quantization for storage, while activations and other sensitive layers remain in full precision. No separate pruned models are hosted on public endpoints; all are original, unpruned versions.

Source: inference-docs.cerebras.ai