Liquid AI has released DSpark draft model checkpoints for three models in its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These drafters add a speculative decoding path to existing target models, allowing a roughly 300M-parameter draft to propose a block of nine candidate tokens that the target model verifies in a single forward pass.

  • The system delivers up to 3.18x faster decoding on an H100 and up to 2.87x on an M4 Max MacBook Pro without changing model outputs.
  • Under greedy decoding, the emitted sequence is identical to the target model running alone, ensuring benchmark accuracy remains unchanged.
  • DSpark combines a DFlash-style parallel backbone with a lightweight sequential head modeled as a Markov chain and a confidence-scheduled verifier.
  • Multi-tool function calling scenarios see an average 57% latency reduction for LFM2.5-2.6B, though MoE models gain only 1.18x on Apple silicon due to current implementation limits.
  • Weights are available as Safetensors and GGUF, with day-one support in llama.cpp and SGLang.

The technology allows developers to significantly reduce inference latency for local coding assistants, on-device agents, and offline copilots while maintaining data privacy through self-hosting.