Liquid AI: 3.18x Faster Decoding with LFM2.5-DSpark

Liquid AI has released new DSpark draft model checkpoints for three of its LFM2.5 models. These new models offer significantly faster decoding speeds without changing the output. What's interesting is that decoding can be up to 3.18 times faster on an H100 and up to 2.87 times faster on an M4 Max MacBook Pro. This speed increase comes from adding a speculative decoding path to existing models. A smaller, 300-million parameter draft model proposes nine candidate tokens, which the larger model then verifies in a single pass. This means the emitted sequence is identical to the original model, so accuracy remains unchanged. The weights are available as Safetensors and GGUF, and day-one support is available for both llama.cpp and SGLang. This technology can be deployed if you self-host. It's suitable for independent developers, startups, and small to medium-sized businesses with under $10 million in annual revenue. The bottom line is that these DSpark draft models offer a way to achieve much faster performance for AI applications without sacrificing accuracy.