Liquid AI Releases DSpark Draft Model for LFM2.5-VL-3B, Claims Up to 3.13x Decode Speedup

Liquid AI has released LFM2.5-VL-3B-DSpark, an experimental draft model that accelerates inference for its vision-language model LFM2.5-VL-3B through speculative decoding. According to Liquid AI, the drafter delivers decode speedups of up to 3.13x on-device and 2.66x on Nvidia H100 GPUs, with end-to-end latency improvements up to 2.62x and 2.27x respectively, without altering output quality.

What DSpark Adds

The drafter adds 280 million parameters to the 3-billion-parameter target model — an 8.9% increase in deployed parameter count. It uses a 4-layer, attention-only architecture with a block size of 8 or 9 tokens, chosen after ablations across 3, 4, and 5-layer configurations. Liquid AI trained the model for 10 epochs on a mixture of vision-language supervised fine-tuning data weighted toward expected production workloads.

Architecturally, the drafter mirrors Liquid AI's earlier text-only DSpark models released in August 2026. It captures the target model's hidden states at fixed tapped layers and conditions on them to draft blocks of candidate tokens. Because image patches and text tokens are projected into a shared representation before those tapped layers, the drafter operates on hidden-state vectors of identical dimensionality regardless of input modality — meaning the underlying speculative decoding algorithm is unchanged from the text version.

Benchmark Results

Liquid AI tested the drafter across six vision tasks — general VQA, text VQA, image captioning, chart VQA, complex reasoning, and multi-turn conversation — following the MMSpec benchmark methodology.

On-device results:

• MLX on Apple M5 Max: decode speedup 2.30x–3.13x; end-to-end speedup 1.56x–2.62x
• llama.cpp on Apple M3 Ultra: decode speedup 1.57x–2.14x; end-to-end speedup 1.30x–1.77x

GPU results:

• H100: decode speedup up to 2.66x; end-to-end speedup 1.64x–2.27x

Speculative decoding is exact — the target model verifies every token proposed by the drafter, so greedy output matches running the target model alone. Liquid AI reports per-response timings via draft_n and draft_n_accepted metrics.

Why Gains Are Smaller for Vision Models

Liquid AI notes a structural limitation: speculative decoding only accelerates the decode phase, not vision encoding or prefill. In VLMs, images pass through a vision encoder before the language backbone processes hundreds of visual tokens alongside text, making prefill a larger share of total latency — especially on edge devices with less compute than datacenter GPUs. The company frames this as an application of Amdahl's law: overall speedup is capped by the portion of the workload that isn't accelerated. Liquid AI notes that Apple's M5 chip narrows this gap somewhat via per-core GPU neural accelerators.

Availability

LFM2.5-VL-3B-DSpark is available now on Hugging Face in Safetensors and GGUF formats. Liquid AI has built day-one support into three inference engines:

• SGLang: requires a build with DSpark support for LFM2 targets (PR #40651)
• llama.cpp: requires PR #29339
• MLX-VLM: requires PR #2280

All three expose the drafter through standard flags (--speculative-draft-model-path, -md, --draft-model) pointed at the target model plus the drafter checkpoint. The model is open-weight, per Liquid AI, with no stated restrictions on fine-tuning or deployment. Pricing not applicable — this is a self-hosted open-weight release, not a hosted API product.

What This Means

Speculative decoding drafters have become a standard technique for squeezing extra throughput out of deployed LLMs without retraining the target model, and Liquid AI is extending that playbook to multimodal models — a harder problem given the added cost of vision encoding and prefill. The reported gains (up to 3.13x decode, up to 2.62x end-to-end) are real speedups if the benchmarks hold up under independent testing, but Liquid AI's own data shows the vision encoder and prefill stages cap how much end-to-end latency improves, particularly on edge hardware. For teams running LFM2.5-VL-3B on-device or on GPU infrastructure, this is a low-cost drop-in accelerator — an 8.9% parameter overhead for a meaningful decode speedup — rather than a new capability. It signals that Liquid AI is treating draft models as a standard companion artifact across its model family, following the same pattern established with its text-only DSpark release last month.