Liquid AI releases speculative decoding drafts for LFM2.5

Liquid AI announced DSpark draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. The company says this is the first public release of speculative decoding models for its Liquid Foundation Models. The checkpoints are on Hugging Face, and LFM-compatible DSpark support has been open-sourced upstream in llama.cpp, with GGUF checkpoints, and in SGLang.

DSpark has three parts. A DFlash-style parallel backbone drafts a block of tokens in one forward pass. A lightweight Markov-chain head conditions each token on the one before it, which Liquid says raises acceptance at later positions. A confidence head and a hardware-aware scheduler then cut off low-confidence suffixes when verifying them would cost more than it saves. Under greedy decoding the output is identical to the baseline by construction, so benchmark accuracy does not change.

Each draft model has about 300M parameters (295.7M for the 1.2B target, 327.7M for the other two) and uses five attention-only layers with a block size of 9. Liquid trained the drafts for 15 epochs, and all training and ablations ran on AMD hardware. For the larger targets, validation loss kept falling after acceptance rates had flattened. Liquid therefore published the epoch with the highest acceptance rate rather than the lowest loss.

The speedups were measured at batch size 1 and temperature 0 on MATH500, GSM8K, HumanEval, MBPP and MT-Bench. Liquid used SGLang on a single H100 in BF16 and llama.cpp with experimental Metal kernels on an M4 Max MacBook Pro in FP16. The post's results for LFM2.5-2.6B were cut off in the source material, so per-benchmark figures are not available here. Performance at larger batch sizes was not reported.