The target model verifies every token proposed by the draft model.
The repository lists 327.7 million BF16 draft parameters, five full-attention layers, a 128,000-token vocabulary, a rank-256 Markov head, a confidence head, and block size 9. Its SGLang example uses LFM2.5-24B-A2B-FP8 as the target and requires a recent build with LFM2-MoE DSpark support.
Liquid AI reports measurements on one NVIDIA B300 with an FP8 target and BF16 KV cache: arithmetic-mean output throughput was 2.38x, 2.69x, and 1.71x versus target-only decoding at concurrency 1, 32, and 256 respectively. The source does not establish task-quality gains, deployment cost, or equivalent results on other hardware. It also notes that floating-point differences can change generated text, including with greedy sampling.