Liquid AI has open-sourced the DSpark draft models for its LFM2.5 series, enabling speculative decoding that delivers up to 3.18x throughput improvement on GPUs and 2.87x on devices.

  • The DSpark architecture combines a parallel backbone, a lightweight sequential head, and a confidence-scheduled verifier to reduce latency in the memory-bound decode phase.
  • Day-one support is provided for llama.cpp and SGLang, allowing users to run the models with minimal configuration changes.
  • On an M4 Max MacBook Pro, LFM2.5-2.6B achieves speeds exceeding 140 tok/s, while function-calling latency is reduced by 57% on average.
  • Draft model checkpoints are available in both Safetensors and GGUF formats for various LFM2.5 sizes including 1.2B, 2.6B, and 8B.

The integration allows users to deploy efficient agentic inference locally or on accelerators without sacrificing the accuracy of greedy decoding.