Up to 3.2x Faster Inference with LFM2.5-DSpark

THE NEWS

What happened

Liquid AI has released LFM2.5-DSpark, a new speculative decoding technique that accelerates Large Language Model inference by up to 3.2x on both high-end GPUs and consumer MacBooks. By combining parallel backbone processing with a lightweight sequential head and a confidence-scheduled verifier, the technology reduces latency significantly without compromising output quality. The solution offers immediate support for popular open-source frameworks like llama.cpp and SGLang, enabling faster local deployment of 1.2B to 8B parameter models.

CONTEXT

Why it matters

Liquid AI releases LFM2.5-DSpark: up to 3.2x faster inference on H100 and MacBooks. Cuts function-calling latency by 57%. Open-source support for llama.cpp and SGLang now available.

AT A GLANCE

Key facts

The main verified points:

• LFM2.5-DSpark achieves up to 3.18x throughput improvement on H100 GPUs and 2.87x on-device performance.
• Function-calling latency is reduced by an average of 57% for the LFM2.5-2.6B model.
• The architecture integrates three components: parallel backbone, sequential head, and confidence-scheduled verifier.
• Draft models are approximately 300M parameters and maintain quality parity with greedy decoding baselines.
• Day-one support is available for llama.cpp and SGLang via open-source upstream integration.
• On-device testing was conducted on an M4 Max MacBook Pro using FP16 GGUF weights.