Liquid AI LFM2.5 DSpark Hits 3.18x

This is a practical inference win for edge and GPU deployments, though the headline speedup depends heavily on workload and token acceptance.

• –Draft models propose multiple tokens while the target verifies them in one forward pass
• –The 8B-A1B model reached 3.18× on MATH500 but only 1.29× on GSM8K
• –Output quality remains unchanged because rejected draft tokens are replaced by the target model
• –Official checkpoints cover three LFM2.5 sizes, with llama.cpp and SGLang deployment paths
• –Developers should benchmark their own prompts before assuming equivalent production gains