TokenSpeed: an open LLM inference engine built for coding agents

The LightSeek Foundation released TokenSpeed, an open-source (MIT) inference engine aiming for TensorRT-LLM-level performance with vLLM-level usability, tuned for agentic workloads with long contexts and a ~70 tokens/sec per-user floor. The team reports that on Kimi K2.5 with NVIDIA B200 it beat TensorRT-LLM by about 9% at batch size 1 and delivered about 11% more throughput.

• Open-source MIT inference engine from the LightSeek Foundation
• Targets TensorRT-LLM performance with vLLM usability
• On Kimi K2.5 and B200: +9% at batch size 1, +11% throughput
• Up to 580 tokens/sec on Qwen3.5-397B-A17B for agentic workloads