Liquid AI Releases DSpark Draft Models for Speculative Decoding on LFM2.5

Liquid AI published DSpark draft model checkpoints for three LFM2.5 models on August 20, 2026. The roughly 300M-parameter drafters deliver up to 3.18x throughput on an H100 and up to 2.87x on-device in the company's own measurements, with output identical under greedy decoding.

On August 20, 2026, Liquid AI released DSpark draft model checkpoints for three models in its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B¹. Each drafter runs around 300M parameters and, in the company’s description, trades a minimal memory increase for a large decoding speedup without changing output quality¹. The figures given are up to 3.18x throughput on a GPU, up to 2.87x on-device, and a 57% average reduction in function-calling latency for LFM2.5-2.6B¹.

Speculative decoding has a lightweight draft model produce candidate tokens that the target model verifies in a single forward pass, sharing the cost of loading weights across all verified tokens¹. Under greedy decoding the target verifies every token, so the emitted sequence is identical to running the target alone, leaving benchmark accuracy unchanged¹. All speed figures are Liquid AI’s own measurements, taken on an M4 Max MacBook Pro with llama.cpp/Metal and a single H100 80GB with SGLang, using a block size of 9, batch size 1, temperature 0, across five benchmark datasets¹. Both llama.cpp and SGLang are supported from day one, though each requires the corresponding build¹. The company also released LFM2.5-VL-3B, a vision-language model with improved screen and UI understanding on August 12².

Sources

Up to 3.2x Faster Inference with LFM2.5-DSpark - Liquid AI blog on Hugging Face (August 20, 2026)
LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge - Liquid AI blog on Hugging Face (August 12, 2026)