Liquid AI has released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model, adding a speculative decoding path that the company says delivers up to 3.13× faster decoding on Apple silicon and 2.66× on NVIDIA H100 GPUs. The drafter adds roughly 280 million parameters — an 8.9% increase over the 3B target — and works out of the box with llama.cpp, MLX-VLM, and SGLang from launch day.

The release extends Liquid AI's DSpark approach from text-only models to vision-language workloads. The vision drafter uses the same architecture as the text LFM2.5-DSpark drafters: at several fixed points in the network, it samples the target model's hidden states and turns those signals into a block of draft tokens. Before those layers, both visual patches and textual inputs are mapped into one common representation, so the drafter always works with hidden-state vectors of the same dimensionality no matter the input type.

Training followed the DSpark recipe with a mixture of vision-language supervised fine-tuning data weighted toward expected serving workloads. Based on ablations across three, four, and five layers, the final draft model is a simplified attention-only drafter with four layers and a block size of nine. The team trained for ten epochs on that mixture, checking acceptance rates at every epoch boundary; gains kept climbing as training tokens increased, but eventually flattened out. At inference time, Liquid AI recommends a block size of eight or nine depending on the hardware.

On-device inference with MLX on an M5 Max shows decoding speedups of 2.30× to 3.13× across six vision tasks from the MMSpec benchmark, with end-to-end latency improvements of 1.56× to 2.62×. Running llama.cpp on an M3 Ultra yields decode gains of 1.57× to 2.14×, while full-pipeline latency improves by 1.30× to 1.77×. On H100 GPUs, the same drafter delivers 2.04× to 2.66× faster decoding, with end-to-end improvements of 1.64× to 2.27×.

The company notes a key limitation: speculative decoding accelerates only the decode stage, not vision encoding or prefill. In a VLM pipeline, visual input is first handled by a vision encoder, after which the language backbone must work through hundreds of visual tokens together with the text prompt. On edge devices, the prefill stage accounts for a larger share of total latency, so even a large decode speedup yields only a modest overall gain — an instance of Amdahl's law where the unaccelerated portion caps total speedup.

The draft model can be downloaded from Hugging Face, with both Safetensors and GGUF weights on offer. Running with SGLang requires a build with DSpark support for LFM2 targets (PR #40651); llama.cpp requires PR #29339; MLX-VLM requires PR #2280. The block size is pulled from the draft model's configuration file or sidecar metadata, and the speculative process stays lossless: the target model checks each drafted token, so greedy results match what the target would produce on its own.

Confirmed

  • LFM2.5-VL-3B-DSpark draft model released September 24, 2026
  • Drafter: ~280M parameters (8.9% overhead on 3B target)
  • Architecture: 4-layer attention-only drafter, block size 9 (recommended 8 or 9 at inference)
  • Training: 10 epochs on vision-language SFT mixture, weighted toward serving workloads
  • Day-one integrations: llama.cpp (PR #29339), MLX-VLM (PR #2280), SGLang (PR #40651)
  • On-device (MLX, M5 Max): decode 2.30–3.13×, end-to-end 1.56–2.62× across 6 MMSpec tasks
  • On-device (llama.cpp, M3 Ultra): decode 1.57–2.14×, end-to-end 1.30–1.77×
  • GPU (H100): decode 2.04–2.66×, end-to-end 1.64–2.27×
  • Formats: Safetensors and GGUF on Hugging Face
  • Speculative decoding is exact; target verifies every token
  • Limitation: prefill and vision encoding not accelerated; Amdahl's law caps end-to-end gains

Unknown

  • Independent replication of vendor-reported speedups across hardware configurations
  • Power consumption and thermal impact of the drafter on edge devices
  • Performance on vision tasks outside the MMSpec benchmark suite
  • Whether the 8.9% parameter overhead translates to proportional memory bandwidth pressure
  • Long-term maintenance and update cadence for the three inference backends

Our take

Liquid AI's vision DSpark release is a pragmatic extension of its text drafter work — same architecture, same integration targets, same exact-decoding guarantee. The speedup numbers are real but bounded by prefill dominance on edge hardware, a constraint the company acknowledges upfront. For developers deploying VLMs on Macs or workstations, the 1.5–2.6× end-to-end gains are meaningful without requiring model surgery. The open-weight release and three backend integrations lower the adoption barrier considerably.

Sources