Liquid AI has released new draft checkpoints called DSpark to make its LFM2.5 models run faster on servers and local devices. If you run AI models locally or build with open-source tools, this update can cut waiting times without changing the final text output.
According to a Liquid AI / Hugging Face post published on 2026-08-20, DSpark adds speculative decoding to three models: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. In this setup, a small draft model predicts upcoming words quickly. Then, the main model checks them in one pass. Under standard greedy decoding, any bad draft token is replaced by the main model. That means the final output remains identical.
The numbers Liquid AI reports
Liquid AI shared strong benchmark numbers across several setups:
- Up to 3.18x higher GPU throughput on a single Nvidia H100 80GB running in BF16 precision.
- Up to 2.87x faster on-device speed on an Apple M4 Max MacBook Pro using FP16 GGUF weights in Metal.
- 57% lower latency on average for function calling with LFM2.5-2.6B.
- 18% average speedup on-device for the MoE model LFM2.5-8B-A1B, where Metal currently activates extra experts during verification.
The integration is ready today. You can already use it via llama.cpp PR #27383 and SGLang PR #31041.
The part you shouldn’t trust blindly
All of these performance numbers come straight from Liquid AI’s internal benchmarks. No independent third party has tested them yet. Liquid AI ran tests with block size 9, batch size 1, temperature 0, up to 256 output tokens, and five benchmark datasets. The results do not guarantee savings on different chips, in large batches, or with creative temperature settings. Liquid AI also noted that speedups on the 1.2B model varied by as much as 52% depending on the text.
Speed is never guaranteed.
What this means for your setup
Speculative decoding is not magic. It only helps if the small model guesses correctly most of the time. If the draft fails, checking the guesses wastes time.
Still, day-one support in llama.cpp and SGLang makes testing easy. Local speed matters.
How to test it yourself
Do not trust averages. Run your own test before changing your production pipeline.
Take the exact DSpark draft and target weights for your LFM2.5 model. Load them in your current llama.cpp or SGLang build. Run your usual prompts and function calls with and without DSpark. Measure tokens per second, memory, draft acceptance rate, and tool-call success. If the speedup holds on your hardware and your tools return valid arguments, DSpark is an easy win.