LFM2.5-DSpark Speeds Local AI Inference Up to 3.2x

Liquid AI released LFM2.5-DSpark on August 20, 2026, and it is a release worth understanding because of what it is not. DSpark is not a new frontier model. It is a set of draft model checkpoints that make the LFM2.5 models you already run go substantially faster, without changing a single output token.

• Three draft checkpoints, each roughly 300M parameters: 295.7M for LFM2.5-1.2B-Instruct, 327.7M for LFM2.5-2.6B, and 327.7M for LFM2.5-8B-A1B
• Up to 3.18x throughput improvement on GPU and up to 2.87x on-device
• Function-calling latency dropped 57% on average for LFM2.5-2.6B
• Available on Hugging Face in Safetensors and GGUF, with day-one support for llama.cpp and SGLang

How Does Speculative Decoding Actually Work?

Speculative decoding is one of the more elegant tricks in local inference. A lightweight draft model proposes several candidate tokens ahead. The larger target model then verifies all of them in a single forward pass. Because loading the big model's weights is the expensive part, verifying five tokens costs barely more than generating one, and any tokens the target accepts come nearly free.

The critical property is that output quality is unchanged. The target model still decides what is correct; the draft model only guesses what the target is likely to say next. When the guess is right you save time. When it is wrong you fall back to normal generation. There is no quality tradeoff to weigh, which is unusual for a speedup this size.

Why 300M Parameters Is the Right Size for a Draft Model

Draft models live or die on a ratio. Too small and the acceptance rate collapses, so you burn compute on rejected guesses. Too large and the drafting itself costs as much as it saves. Landing all three checkpoints near 300M parameters against targets from 1.2B to 8B keeps the drafting overhead low while staying close enough to the target's distribution to be accepted often.

The function-calling result is the one that will matter most in practice. Tool-calling agents generate short, highly structured outputs where a draft model can predict most of the JSON scaffolding correctly, and cutting that latency by 57% on average is the difference between an agent that feels responsive and one that feels like it is thinking. Our earlier look at LFM2.5-2.6B running tool-calling agents in 2.5GB of RAM covers the target model that benefits most here.

What This Means for Self-Hosted LLM Hardware

Speculative decoding shifts where the bottleneck sits. On memory-bandwidth-limited hardware, which describes almost every mini PC and laptop running a local LLM, the win is largest, because bandwidth is exactly what verifying a batch of draft tokens amortises. That is why the on-device figure of up to 2.87x is arguably more interesting than the 3.18x GPU number.

Day-one llama.cpp and GGUF support means this reaches self-hosters immediately rather than after a porting cycle. For anyone tracking the broader local AI and open weights beat, DSpark continues a pattern Liquid AI has been running all year: instead of chasing benchmark headlines, ship the unglamorous pieces that make small models practical, as they did with LFM2.5 vision running screen understanding in 3GB.

One Honest Caveat on the 3.2x Figure

Speculative decoding speedups are not a flat multiplier. They track how often the target model accepts the draft's guesses, and that acceptance rate varies by task rather than by model size. Liquid AI's own figures make the point: LFM2.5-8B-A1B accepts 8.27 of every 10 proposed tokens per step on MATH500, which produces the headline 3.18x, but only 4.02 out of 10 on GSM8K, which lands nearer 1.29x on identical hardware.

That is a feature of the technique, not a flaw in this implementation, and it is worth knowing before you benchmark your own workload. Structured, predictable output accelerates most. Open-ended reasoning accelerates least. Liquid AI is also the first to publish DSpark-compatible draft checkpoints as open weights for a separate model family, which lowers the barrier for anyone who wants to test the ratio on their own tasks.

Sources: Liquid AI — August 20, 2026; Hugging Face — August 20, 2026; MarkTechPost — August 20, 2026.