Liquid AI has released draft model checkpoints for its LFM2.5 family—specifically LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B—that utilize speculative decoding to significantly improve inference speed without compromising output quality.

  • The DSpark method combines a parallel backbone, a lightweight sequential head, and a confidence-scheduled verifier to reduce latency by trading minimal memory for large throughput gains.
  • Benchmarks show up to 3.18x throughput improvement on NVIDIA H100 GPUs and up to 2.87x on edge devices like the M4 Max MacBook Pro.
  • For agentic workloads involving function calling, DSpark reduces latency by an average of 57% on the BFCL dataset.
  • The draft models are open-sourced upstream in llama.cpp and SGLang, with checkpoints available on Hugging Face in both Safetensors and GGUF formats.

This release aims to make LFM2.5 viable for on-device agentic applications by providing day-one support for efficient edge inference and high-throughput GPU serving.