Hugging Face publishes DSpark speculative decoding draft models for Liquid AI’s LFM2.5 family

Liquid AI published three DSpark speculative decoding draft model checkpoints for its LFM2.5 family on Hugging Face: LFM2.5-1.2B-Instruct-DSpark, LFM2.5-2.6B-DSpark, and LFM2.5-8B-A1B-DSpark. The draft models are provided in Safetensors and GGUF formats and integrate with llama.cpp and SGLang for accelerated inference.

The DSpark approach uses a lightweight draft model to generate candidate tokens and a target model to verify them in a single forward pass, sharing the cost of loading weights across verified tokens. Draft models are around 296–328M parameters, with a 5-layer attention-only backbone, a hidden-state projection layer, a Markov head for inter-token dependency, and a confidence-scheduled verifier to prune low-confidence suffixes.

Across five benchmark datasets, the draft models deliver measurable throughput improvements on both GPU and on-device configurations. On an H100 80 GB, LFM2.5-2.6B-DSpark achieves a mean 2.67× speedup (up to 3.06× on MATH500). On an M4 Max MacBook Pro with FP16 GGUF weights, the same model achieves a mean 2.27× speedup (up to 2.63× on HumanEval).

For LFM2.5-1.2B-Instruct-DSpark, mean speedups are 2.10× on H100 and 2.54× on M4 Max. For LFM2.5-8B-A1B-DSpark, mean speedups are 2.54× on H100 and 1.18× on M4 Max, with the on-device gap attributed to the current MoE implementation in llama.cpp’s Metal backend.

The draft models reduce function-calling latency by 57% on average for LFM2.5-2.6B across various multi-tool scenarios. Output quality is preserved because the target model verifies every proposed token under greedy decoding, ensuring identical sequences to the baseline.

To use the draft models with SGLang, an SGLang build with DSpark support for LFM targets is required; the server is launched with the draft attached via command-line flags. With llama.cpp, a compatible build is needed and the draft is specified via sidecar metadata.

The checkpoints are available on Hugging Face under Liquid AI’s license and include both Safetensors and GGUF formats for the three draft models.