Issue 2026-11-14 · Industry · 开源 · 模型发布 · 芯片
DeepSeek Ultrafast engine open-sourced with 1M context

The open-source DeepSeek Ultrafast inference engine for DeepSeek-V4.1-Flash supports 1M-token context and multi-server KV cache sharing. On eight B300 GPUs it reaches 2,060 tokens/s decode, 2.71 ms step time and 11.7 s to prefill 1M tokens. The model is available via Baseten.
Read original ↗