Liquid AI launches LFM2.5-DSpark, boosting model speed up to 3.2x
Liquid AI releases the LFM2.5-DSpark drafter model suite, increasing inference speed by up to 3.2x without compromising quality.
Liquid AI has announced the launch of LFM2.5-DSpark, a "drafter" model suite designed to work with Speculative Decoding techniques to accelerate AI processing and inference for the LFM2.5 model family while maintaining output quality.
According to the released details, LFM2.5-DSpark delivers an inference speedup of up to approximately 3.2x (specifically 3.18x), supporting both high-performance data center GPUs (such as the H100) and edge devices like MacBooks.
For this release, the developer has rolled out three models to support the main models in the family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. Additionally, the tool has been designed for seamless integration with popular LLM serving frameworks such as SGLang and llama.cpp.
Enables developers and users to run LFM2.5 models significantly faster across both high-end hardware and consumer devices, reducing AI processing time and costs.