Liquid AI ships a 280M helper to speed up its 3B vision model

reported by 2 outlets· 2 articles · confidence: high · first seen 2026-09-24 00:00 UTC

What this means for you

If you serve LFM2.5-VL-3B, this is close to a drop-in: the drafter weights sit alongside the target and llama.cpp, MLX-VLM and SGLang support it from day one, for 8.9% more parameters. Benchmark it at your own batch size — the reported gains are largest in decode-bound, low-batch settings.

Liquid AI has published a 279.5M-parameter companion model that makes its LFM2.5-VL-3B vision-language model decode faster. The method is speculative decoding: a small drafter proposes several tokens and the larger model checks them in one pass, so fewer passes are needed. Liquid AI reports 2.30x to 3.13x faster decoding with MLX on an M5 Max, 1.57x to 2.14x with llama.cpp on an M3 Ultra, and up to 2.66x on an H100. Output is unchanged under greedy decoding. The drafter adds 8.9% to the parameter count and ships with day-one llama.cpp, MLX-VLM and SGLang support. One figure in Liquid AI's post, a 20.4x H100 lower bound, appears to be a typo.

Key facts

  • ·Liquid AI released LFM2.5-VL-3B-DSpark, a 279.5M-parameter draft model for its LFM2.5-VL-3B vision-language model. source
  • ·Decoding runs 2.30x to 3.13x faster with MLX on an M5 Max, and 1.57x to 2.14x with llama.cpp on an M3 Ultra, with end-to-end gains of 1.56x to 2.62x and 1.30x to 1.77x respectively. source
  • ·On an H100 the same drafter gives up to 2.66x faster decoding and up to 2.27x end-to-end improvement. source
  • ·The drafter adds roughly 280M parameters, an 8.9% increase on top of the 3B target model. source
  • ·Day-one integration ships for llama.cpp, MLX-VLM and SGLang. source
  • ·The drafter is a 4-layer attention-only model trained at block size 9; Liquid AI recommends block size 8 or 9 at inference, and measurements cover six vision tasks following the MMSpec benchmark. source

What the sources say

  • Hugging Face Blog — The primary technical post: drafter architecture, training mixture, block size and per-hardware speed tables.
  • MarkTechPost — Short news write-up confirming the release, parameter count, hardware ceilings and runtime support.

Sources

The original reporting. Follow these — they did the work.

← the wire