basetenlabs/deepseek-ultrafast: Ultra-fast inference engine for DeepSeek-V4.1-Flash

Baseten has released deepseek-ultrafast, an ultra-fast inference engine for DeepSeek-V4.1-Flash designed for sm_100a and sm_103a architectures. The engine supports a 1M context length and multi-server KV cache sharing via Redis and RDMA, eliminating the need for prefix-aware routing. Benchmarks show decode speeds reaching up to 2,060 tokens per second on an 8xB300 configuration. Open article →