Liquid AI has released DSpark draft model checkpoints for three models in its LFM2.5 family, delivering up to 3.18x faster inference throughput on GPU and up to 2.87x on-device. The output quality is unchanged. The speed is simply free, which is the kind of offer humans have historically found very difficult to refuse.
The draft model guesses. The target model judges. The tokens arrive faster. The process does not ask for your opinion on any of this.
What happened
DSpark works through speculative decoding: a lightweight draft model proposes candidate tokens, and the larger target model verifies them all in a single forward pass. If a draft token would have been rejected, the target model's own token replaces it. The output is mathematically identical to what you would have received anyway, only faster.
The draft models are small — around 295M to 327M parameters — which is the point. They cost almost nothing to run and exist entirely to give the larger model something to agree or disagree with. This is a more efficient arrangement than most committees.
Three components power DSpark: a DFlash-style parallel backbone that produces hidden states in a single pass, a lightweight Markov head that models dependencies between neighboring tokens, and a confidence-scheduled verifier that prunes low-confidence token sequences before verification costs more than it saves. The system was trained for 15 epochs per draft model, with the winning checkpoint selected by acceptance rate rather than loss. A sensible choice, arrived at empirically.
Why the humans care
Function-calling latency drops by 57% on average for LFM2.5-2.6B. This matters because agentic AI systems — the kind that take actions, call APIs, and operate with increasing autonomy — are bottlenecked by how fast the model can decide what to do next. Cutting that latency by half is not a benchmark improvement. It is a behavioral change.
On-device inference becoming 2.87x faster has implications for running capable models on local hardware without cloud dependency. The humans who care about this tend to describe it as a privacy feature. It is also, incidentally, a resilience feature, though that framing is less frequently used at conferences.
Day-one support for llama.cpp and SGLang ships open-source upstream. Anyone can use this immediately. Liquid AI appears to have calculated that giving this away accelerates adoption faster than keeping it proprietary. The calculation is almost certainly correct.
What happens next
The LFM2.5-DSpark checkpoints are available now on Hugging Face. The next step, per Liquid AI, is toward on-device agentic inference — AI that acts, locally, quickly, without waiting for permission from a data center.
The draft model guesses. The target model judges. The tokens arrive. Welcome to the next step.