LFM2.5-VL-3B-DSpark (Liquid AI)

What it claims

This is a speculative-decoding draft model for the LFM2.5-VL-3B vision-language model, not a new VLM itself. The drafter adds 280M parameters (8.9% on top of the 3B target) and delivers decode speedups up to 3.13x on device and 2.66x on an H100, with end-to-end gains up to 2.62x (device) and 2.27x (H100). Because verification is exact, greedy output is identical to running the target alone. Note the existing catalog summary treats this as a standalone 3B VLM; it is actually a drafter you attach to LFM2.5-VL-3B.

How it works

The vision drafter reuses the text LFM2.5-DSpark architecture: it taps the target model's hidden states at a fixed set of layers and drafts a block of k candidate tokens the target then verifies. Image patches and text tokens are projected into a shared representation before the tapped layers, so the drafter sees hidden-state vectors of the same dimensionality regardless of modality — meaning the inference algorithm is unchanged from the text case. Training followed the DSpark recipe on a vision-language SFT mixture, 10 epochs. Ablations across 3/4/5 layers settled on an attention-only drafter, 4 layers, block size 9; recommended inference block size is 8–9.

Evidence

Measured on six vision tasks (general VQA, text VQA, captioning, chart VQA, complex reasoning, multi-turn) following the MMSpec benchmark, block size 8:

• MLX on M5 Max: decode 2.30x–3.13x, end-to-end 1.56x–2.62x.
• llama.cpp on M3 Ultra: decode 1.57x–2.14x, end-to-end 1.30x–1.77x.
• H100: decode up to 2.66x, end-to-end 1.64x–2.27x. (The page's "20.4x" lower bound appears to be a typo.)

How to use it

Open-weight, on Hugging Face in Safetensors and GGUF. SGLang (needs DSpark support for LFM2, PR #40651):

python -m sglang.launch_server \
  --model-path LiquidAI/LFM2.5-VL-3B \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path LiquidAI/LFM2.5-VL-3B-DSpark \
  --speculative-draft-attention-backend flashinfer \
  --speculative-dspark-block-size 9 \
  --disable-radix-cache

llama.cpp (PR #29339):

llama-server -m models/LFM2.5-VL-3B-F16.gguf \
  --mmproj models/mmproj-LFM2.5-VL-3B-F16.gguf \
  -md LFM2.5-2.6B-DSpark-F16.gguf \
  --spec-type draft-dspark --spec-draft-n-max 8 --spec-draft-n-min 0 \
  -fa on -ngl 99 -c 8192

MLX-VLM (PR #2280):

mlx_vlm.server --model LiquidAI/LFM2.5-VL-3B --draft-model LiquidAI/LFM2.5-VL-3B-DSpark

Limits

Speculative decoding accelerates only decode, not vision encoding or prefill. On edge devices, prefill (vision encoder + hundreds of visual tokens) dominates wall time, so Amdahl's law caps end-to-end gains well below the decode speedup — this is why device end-to-end numbers trail the decode figures. Each backend requires a specific unmerged/PR build, so day-one support is gated on those PRs landing in your toolchain.