Cerebras has announced the availability of the Qwen 3.8 27B model on its platform, achieving an inference throughput of 1500 tokens per second. To maintain high fidelity, Cerebras serves unpruned, original versions of its hosted open-source models through public endpoints, ensuring the architecture remains unmodified. While the company conducts research into Router-weighted Expert Activation Pruning (REAP)—with these specialized models released via Hugging Face for experimental use—production models are not subjected to pruning. Regarding quantization, Cerebras employs a selective weight-only strategy during storage to balance efficiency with model quality. Specifically, weights are stored in partial 16-bit, 8-bit, or 4-bit formats, but sensitive layers are preserved at full precision and undergo on-the-fly dequantization to ensure high-precision operations. Crucially, all activations, attention mechanisms, and KV caches are maintained in a full-precision, unquantized state to prevent performance degradation. This methodology distinguishes their production API from their research collection, where pruned models are offered as separate entities to maintain transparency for developers choosing between raw model quality and optimized architectural versions.