---
title: "Up to 3.2x Faster Inference with LFM2.5-DSpark"
description: "Liquid AI released DSpark draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B, adding speculative decoding with a minimal memory increase. DSpark uses a DFlash-style parallel backbone conditioned on target context features, a lightweight sequential Markov-chain head, and a confidence-scheduled verifier that prunes low-confidence suffixes. The draft models are simplified attention-only networks with five layers and a block of nine, trained for 15 epochs on SFT, chat, code, and function-calling data; each contains roughly 300M parameters. Across five datasets, LFM2.5-2.6B averaged 2.67x speedup on an H100 and 2.27x on an M4 Max, while reducing multi-tool function-calling latency by 57% on average. Greedy decoding remains identical to baseline by construction, and integrations are available in llama.cpp and SGLang, with on-device gains for the 8B-A1B model limited by its current MoE Metal implementation."
---

# Up to 3.2x Faster Inference with LFM2.5-DSpark

[Hugging Face](https://yomu.fyi/company/hugging-face) · xx, Leonie Monigatti, Fernando Fernandes Neto, Tarek Dakhran, nathan ranchin · Aug 20, 2026

**Type:** Benchmark

## Summary

Liquid AI released DSpark draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B, adding speculative decoding with a minimal memory increase. DSpark uses a DFlash-style parallel backbone conditioned on target context features, a lightweight sequential Markov-chain head, and a confidence-scheduled verifier that prunes low-confidence suffixes. The draft models are simplified attention-only networks with five layers and a block of nine, trained for 15 epochs on SFT, chat, code, and function-calling data; each contains roughly 300M parameters. Across five datasets, LFM2.5-2.6B averaged 2.67x speedup on an H100 and 2.27x on an M4 Max, while reducing multi-tool function-calling latency by 57% on average. Greedy decoding remains identical to baseline by construction, and integrations are available in llama.cpp and SGLang, with on-device gains for the 8B-A1B model limited by its current MoE Metal implementation.

## Context

The decode phase in LLM inference is traditionally memory-bound because most latency comes from streaming weights from DRAM into SRAM rather than from computation. The work also targets lower latency for on-device agentic inference, including function-calling scenarios.

## Approach / What changed

DSpark uses speculative decoding: a lightweight draft model proposes candidate tokens, and the target model verifies them in one forward pass. Its design combines a DFlash-style parallel backbone, a sequential Markov-chain head, and a confidence-scheduled verifier. The released draft models are simplified five-layer attention-only models trained on SFT, chat, code, and function-calling data, with integrations for llama.cpp and SGLang.

## Takeaways

- LFM2.5-2.6B averaged 2.67x speedup on an H100 and 2.27x on an M4 Max across MATH500, HumanEval, MBPP, GSM8K, and MT-Bench.
- Greedy speculative decoding preserves baseline output exactly because rejected draft tokens are replaced by the target model's own token.
- The LFM2.5-8B-A1B draft model averaged only an 18% on-device improvement because verifying multiple tokens activates more experts and increases weight traffic in the current Metal MoE implementation.

**Tags:** [Machine Learning](https://yomu.fyi/topic/machine-learning), [Open Source](https://yomu.fyi/topic/open-source), [Performance](https://yomu.fyi/topic/performance)

- Source: [Hugging Face](https://huggingface.co/blog/LiquidAI/lfm25-dspark)
- Source URL: https://huggingface.co/blog/LiquidAI/lfm25-dspark
- Ingested by Yomu: 2026-08-29T05:41:44.083Z

[Read original post](https://huggingface.co/blog/LiquidAI/lfm25-dspark)