Liquid AI Applies DSpark to LFM2.5, More Than Doubling Speed on H100 and MacBook
The DSpark-applied LFM2.5 models show speculative decoding can push compact on-device models past a 2x speedup without trading accuracy, which Liquid AI frames as a step toward running Fable-level intelligence on smartphones.
Reporting from 1 source: GIGAZINE.
Liquid AI applied the speculative decoding technology DSpark to three models in its compact LFM2.5 series: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. The DSpark-applied models achieved more than a 2x speedup while maintaining performance. LFM2.5-2.6B reached a 2.67x speedup on H100 and 2.27x on a MacBook Pro with M4 Max.
Liquid AI built separate draft models for each of the three DSpark versions. The draft model for LFM2.5-1.2B-Instruct has 295.7 million parameters, while LFM2.5-2.6B and LFM2.5-8B-A1B each use a 327.7 million parameter draft model.
On the MacBook Pro with M4 Max, LFM2.5-1.2B-Instruct achieved a 2.54x speedup and LFM2.5-8B-A1B reached 1.18x. The DSpark-applied models are distributed on Hugging Face under the LiquidAI organization.
Liquid AI's Piotr Mazurek called speculative decoding a key part of compressing Fable-level intelligence to run on smartphones, describing this result as a first step toward future performance gains.
- LFM2.5-1.2B-Instruct-DSpark: 2.54x speedup on MacBook Pro with M4 Max
- LFM2.5-2.6B-DSpark: 2.67x speedup on H100 and 2.27x on MacBook Pro with M4 Max
- LFM2.5-8B-A1B-DSpark: 1.18x speedup on MacBook Pro with M4 Max
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.