Liquid AI's LFM2.5-DSpark achieves up to 3.2x faster inference than the base LFM2.5 model on Apple Silicon hardware, according to the Hugging Face blog — with no reported degradation in output quality.
Key takeaways
LFM2.5-DSpark runs up to 3.2x faster than standard LFM2.5 on Apple Silicon, with no reported quality loss.
DSpark is a speculative-decoding variant: a smaller draft model generates token candidates that the larger model verifies in parallel, cutting wall-clock latency.
The gains are hardware-specific: benchmarks focus on Apple Silicon (M-series Macs); results on CUDA GPUs differ.
The model is available on Hugging Face and targets developers and creators running inference locally, not just cloud API users.
Benchmark data shows latency reductions across standard tasks including BFCL function-calling tests, where DSpark posts substantially lower response times.
Liquid AI's LFM2.5-DSpark, the speculative-decoding variant of LFM2.5, released on Hugging Face.
How speculative decoding cuts latency without touching quality
The speed gain is not a compression trick. DSpark uses speculative decoding: a compact draft model proposes a sequence of tokens, and the full LFM2.5 model verifies them in a single forward pass. When the draft is right — which it is, most of the time — the system skips the slow token-by-token generation loop. The result is fewer sequential steps to produce the same output, which is why quality stays intact while wall-clock time collapses.
For anyone running image-captioning pipelines, prompt-expansion workflows, or character-dialogue generation locally on an M-series Mac, that gap is not academic. A 3.2x speedup on a task that previously took eight seconds now takes under three. Iteration loops that felt sluggish become responsive enough to stay in creative flow.
BFCL latency benchmarks show LFM2.5-DSpark posting substantially lower response times than the base model on Apple Silicon hardware.
The benchmark figures are tied to Apple Silicon. M-series chips handle the memory bandwidth demands of speculative decoding particularly well — the unified memory architecture means the draft and verifier models share the same pool without the PCIe bottleneck that can blunt gains on discrete GPU setups. Creators running CUDA-based rigs should treat the 3.2x figure as an upper bound, not a guarantee; Liquid AI's published data does not claim equivalent gains on NVIDIA hardware.
That Apple-first framing is itself a signal. A growing slice of the local-inference community runs on MacBook Pros and Mac Studios — machines that have historically lagged behind NVIDIA in raw throughput but now host a serious ecosystem of quantized and optimized models. DSpark is a direct play for that audience.
What it changes for local creative pipelines
The practical upside reaches beyond raw speed. Faster token generation means prompt-refinement cycles — the back-and-forth of adjusting a character description or scene brief before feeding it to an image generator — compress from minutes to seconds. Creators who use LLMs as a first-pass layer before hitting an image model in their AI image generation workflow will feel the difference immediately.
Speculative decoding also tends to preserve the model's stylistic register because the verifier, not the draft, has the final word on every token. The draft model can be wrong; the large model corrects it. That asymmetry is why Liquid AI can credibly claim no quality loss — the architecture enforces it structurally, not through post-hoc tuning.
Open-weight inference optimization has been accelerating sharply this year, with smaller labs finding headroom that closed-API providers rarely publish. Exploring the model catalog for locally runnable options is increasingly viable for creators who want speed without a cloud bill. DSpark is one of the cleaner examples of that trend: a genuine engineering improvement, available now, with benchmarks attached.
