Up to 3.2x Faster Inference with LFM2.5-DSpark
huggingface ·
Liquid AI released DSpark draft model checkpoints for three LFM2.5 models (1.2B-Instruct, 2.6B, and 8B-A1B), adding a speculative decoding path that speeds up inference with minimal memory overhead. These ~300M-parameter draft models combine a DFlash-style parallel backbone, a lightweight sequential Markov head, and a confidence-scheduled verifier. The approach delivers up to 3.18x throughput improvement on an H100 GPU and up to 2.87x on an M4 Max MacBook, while cutting function-calling latency by 57% on average for LFM2.5-2.6B. Because draft tokens are only accepted when they match the target model's distribution under greedy decoding, output quality and benchmark accuracy remain unchanged. Day-one support is available through llama.cpp and SGLang, with tests using a block size of 9, batch size 1, and temperature 0. The checkpoints are published on Hugging Face in both Safetensors and GGUF formats.
huggingface