Liquid AI's New Draft Models Make Its LFMs Up to 3.2x Faster, No Quality Trade-off
A technique called speculative decoding lets a small helper model do the heavy lifting so the main model just checks the work. The result: dramatically faster output on everything from a data-centre GPU to a MacBook.
Key points
- Liquid AI released DSpark draft models for three LFM2.5 language models in August 2026, cutting inference time by up to 3.18 times on a professional GPU.
- The speed gains come from speculative decoding, a technique where a small helper model proposes tokens and the main model only has to approve or reject them.
- On a MacBook Pro with an M4 Max chip, the LFM2.5-2.6B model averaged roughly 139 tokens per second, faster than most cloud-based AI services.
- Function-calling latency, the delay when an AI model triggers an external tool like a calendar or database, dropped 57% on average for LFM2.5-2.6B.
- Output quality is mathematically identical to the original model: the text produced is word-for-word the same as before.
Liquid AI has released a set of draft models designed to speed up its LFM2.5 family of large language models, the technology that powers AI assistants and coding tools, without changing a single word of what those models produce. The technique is called speculative decoding. A small, cheap helper model guesses ahead several tokens at a time, and the main model checks all those guesses in one quick pass instead of grinding through each token one by one.
Think of it like a fast typist drafting a paragraph for a careful editor. The editor still approves every sentence, so quality never slips, but the pair finishes far faster than the editor working alone. We covered Liquid AI's 2.6B model launch on 4 August, when the base model was shown competing with models four times its size; DSpark is the speed layer added on top.
What does the speed improvement actually look like?
On a single H100 GPU, the kind of specialised chip data centres use for heavy AI work, the LFM2.5-8B-A1B model reached a peak of 3.18 times its previous throughput on maths problems. For conversation tasks, that chip delivered a 3.02-times improvement. On a MacBook Pro with Apple's M4 Max processor, the LFM2.5-1.2B-Instruct model hit 2.87 times its old speed on code-generation benchmarks.
| Model | Best GPU speedup | Best on-device speedup |
|---|---|---|
| LFM2.5-8B-A1B | 3.18x (MATH500) | 1.44x (GSM8K) |
| LFM2.5-2.6B | 3.06x (MATH500) | 2.63x (HumanEval) |
| LFM2.5-1.2B-Instruct | 2.56x (MATH500) | 2.87x (HumanEval) |
The 8B model is an exception on-device: it averaged only an 18% gain on the MacBook. Liquid AI says that's a current limitation of the software layer used to run that particular model type locally, not the technique itself.
Does this change anything for ordinary users?
Responses arrive noticeably sooner if you run any tool built on these models. Nothing else changes. Because every draft token is verified by the full model before it appears, the output is bit-for-bit identical to what you'd have received before.
The sharper gain is for people who connect AI to other software. Function-calling latency, the pause you feel when an AI model reaches out to fetch data or trigger a tool, fell 57% on average for the mid-sized 2.6B model. That matters for any AI "agent", meaning software set up to carry out multi-step tasks on its own, such as searching a calendar or writing to a spreadsheet.
What happens next?
The draft model checkpoints are on Hugging Face in Safetensors and GGUF formats, covering all three model sizes. Support for llama.cpp (for running models locally) and SGLang (for server deployments) shipped on day one. Developers can attach a draft model to an existing LFM2.5 deployment with a handful of extra command-line flags.
For users who don't run their own AI infrastructure, the practical effect arrives when app developers and cloud providers update their deployments. Faster inference typically means lower costs, and lower costs tend to mean faster, cheaper products downstream. The speed story here isn't complete until those updates ship.
Common questions
Does the output change at all?
No. The math behind speculative decoding guarantees that any draft token the main model rejects gets replaced by exactly the token the main model would have chosen anyway.
Can I run this on a regular laptop?
The smaller models, particularly LFM2.5-2.6B, are designed for on-device use. Liquid AI tested them on an M4 Max MacBook Pro, where they ran faster than many cloud services. An older machine will still work but will be slower.
Do I need to be a developer to benefit?
Not eventually. Right now, taking advantage of DSpark requires installing software and running commands. Apps and services built on these models will quietly incorporate the speed gains over time, and end users will simply notice things feel snappier.
