TokenSpeed ships Day-0 DeepSeek V4.1 support, fastest engine on Hopper GPUs
zhyncs42 · x · 2026-10-17
TokenSpeed, an open-source MIT-licensed LLM inference engine (2.7k GitHub stars), announced Day-0 support for DeepSeek V4.1 Flash 1017, claiming it is the fastest DeepSeek V4.1 engine on Hopper GPUs.
Key points:
- Out-of-the-box support for Hopper, Blackwell, Grace Blackwell, and AMD MI355X
- Runs in production across thousands of Hopper GPUs with PD (prefill/decode) disaggregation
- Optimizations for Vera Rubin and AMD MI455X are underway
Related event: TokenSpeed Engine Adds Day-0 Support for DeepSeek V4.1(2 posts)→
More from Infra
- mlsweep: open-source sweep scheduler and visualizer for distributed ML training — burny_tech · 2026-10-17
- Ethereum Foundation's zkAPI brings anonymous frontier AI access via zero-knowledge proofs — DavideCrapis · 2026-10-17
- Goldman: $738B memory stock wipeout after CXMT 4F² reveal was overreaction — zephyr_z9 · 2026-10-17
- Qwen3.6 35B MoE hits 19.5 tok/s on a 16GB 4060 Ti, 10x faster than dense models — dreamyrhodes · 2026-10-17
- Ben Horowitz: a16z avoids neoclouds — bandwidth went from $1,500 to $50/Mbps and wiped out holders — SumitGup · 2026-10-17
- ChinaTalk: Saif Khan argues DUV controls could leave China with 100x fewer AI chips by mid-2030s — ChinaTalk · 2026-10-17