LiquidAI's 280M drafter speeds its 3B vision model by up to 3.13x

single source· 1 articles · confidence: high · first seen 2026-09-24 00:00 UTC

What this means for you

If you serve LFM2.5-VL-3B through llama.cpp, MLX-VLM or SGLang, the drafter is available now, at a recommended block size of 8 or 9. Budget 280M extra parameters, 8.9% of the total, and expect the gain in decoding rather than end-to-end: on an M5 Max, 3.13x decoding against 2.62x end-to-end at best.

LiquidAI has published a drafter for LFM2.5-VL-3B, its 3B vision-language model. A drafter proposes candidate tokens for the larger model to check, so more than one token can be emitted per pass — speculative decoding. It adds 280M parameters, 8.9% on top of the target, and is reported to give decoding speedups of up to 3.13x on an M5 Max, 2.14x on an M3 Ultra and 2.66x on an H100, with end-to-end gains of 1.30x to 2.62x. Support ships for llama.cpp, MLX-VLM and SGLang. No evaluation date is given, and the piece's H100 decoding range starts at 20.4x.

Key facts

  • ·The LFM2.5-VL-DSpark drafter adds 279.5M parameters, 8.9% on top of the 3B LFM2.5-VL-3B target model. source
  • ·Reported decoding speedups are 2.30x to 3.13x with MLX on an M5 Max and 1.57x to 2.14x with llama.cpp on an M3 Ultra; the H100 figure is stated as up to 2.66x in the summary and as a range starting at 20.4x in the body. source
  • ·End-to-end latency gains are reported as 1.56x to 2.62x on MLX/M5 Max, 1.30x to 1.77x on llama.cpp/M3 Ultra and 1.64x to 2.27x on H100. source
  • ·Day-one integrations ship for llama.cpp, MLX-VLM and SGLang. source
  • ·The drafter is a 4-layer attention-only model trained at block size 9; LiquidAI recommends block size 8 or 9 at inference. source
  • ·Evaluation covers six vision tasks — general VQA, text VQA, image captioning, chart VQA, complex reasoning and multi-turn conversation — following the MMSpec benchmark, with no evaluation date stated. source

What the sources say

  • Hugging Face Blog — LiquidAI's own write-up of the drafter's architecture, training recipe and per-device speed measurements.

Sources

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

← the wire