Liquid AI adds a 280M decoding sidecar

Liquid AI released LFM2.5-VL-DSpark, a 280M-parameter draft sidecar that adds 8.9% to its 3B vision-language target and reports up to 2.62x faster end-to-end inference on an M5 Max, according to Liquid AI’s DSpark release notes. Because speculative decoding accelerates decoding rather than image encoding or prompt prefill, the result is a local-inference optimization rather than a universal 2.62x latency cut; edge builders should benchmark their own image-heavy mix beside the memory constraints shaping local AI deployment.