Liquid AI Releases LFM2.5-VL-3B-DSpark, a 279.5M-Parameter Speculative Decoding Drafter for Its Vision-Language Model

The drafter adds about 280 million parameters to the target model. Speculative decoding works by having a small drafter propose several tokens ahead while the larger target model verifies the whole block in a single forward pass and keeps the tokens it agrees with, in place of the one-token-per-pass generation used by a standard model.

Liquid AI's DSpark drafter follows the recipe described in its DSpark paper for text models. It reads the target model's hidden states from several layers and predicts the next k tokens. The company notes that modality does not matter to the drafter: by the time tokens reach the hidden layers, text and image patches are both just tensors, so the same inference algorithm is reused for the vision-language model.

The drafter is a simplified attention-only model. Ablations settled on four layers and a block size of nine, with a recommended inference block size of eight or nine depending on hardware; Apple silicon runs use eight. Its parts are a 193.0-million-parameter decoder stack, a 21.0-million-parameter hidden-state projection, a 65.5-million-parameter Markov head, and a norms and confidence head of about 6,400 parameters, for a total of 279.5 million. The embedding and LM head are tied to the target, so the drafter does not carry them; Liquid AI says this raises the deployed parameter count by 8.9 percent. Training used supervised fine-tuning data covering common vision-language tasks for ten epochs, and all ablation and training work ran on AMD hardware.

Evaluation followed the MMSpec benchmark across six task types: general VQA, text VQA, image captioning, chart VQA, complex reasoning and multi-turn conversation. All runs used batch size 1, temperature 0 and 16-bit weights for the vision encoder and backbone, with data collected on Pipette, Liquid AI's public device-benchmarking infrastructure.

On MLX-VLM with an M5 Max MacBook Pro at block size 8, decode speedups ranged from 2.30x to 3.13x and end-to-end speedups from 1.56x to 2.62x, with 3.24 to 4.34 accepted tokens per pass. On llama.cpp with an M3 Ultra, the ranges were 1.57x to 2.14x and 1.30x to 1.77x, with 3.31 to 4.50 accepted tokens. On SGLang with one H100 80GB at block size 9, they were 2.04x to 2.66x and 1.64x to 2.27x, with 3.46 to 4.57 accepted tokens.

Liquid AI states that the "up to" decode and end-to-end figures often come from different tasks: on the M5 Max, the 3.13x decode result is from COCO captioning while the 2.62x end-to-end result is from MMMU-Pro. Acceptance landed in a similar range on both Apple stacks, which Liquid AI reads as acceptance depending on the drafter and workload rather than the runtime. At higher concurrency, DSpark kept a throughput advantage at every measured level on a single H100 in SGLang, although the gap narrows as concurrency rises.

Under greedy decoding the target verifies every proposed token, so output is identical to the base model, and at non-zero temperatures with matched sampling, speculative decoding preserves the target's output distribution, as proven by Leviathan et al. Temperature does affect speed: higher temperatures spread probability across more candidate tokens, so drafter and target disagree more often, which in Liquid AI's tests lowered acceptance and throughput.

Speculative decoding accelerates decoding only, leaving image encoding and prefill at the same speed. A vision-language model must encode the image and then process hundreds of visual tokens alongside the prompt, and on edge devices with less compute than data center GPUs, prefill takes a larger share of latency. Liquid AI frames this as Amdahl's law, where total speedup is bounded by the part left unaccelerated. This explains cases such as TextVQA on the M5 Max, where 2.69x faster decoding yields 1.56x end to end.

SGLang requires v0.5.19 or newer; users launch LiquidAI/LFM2.5-VL-3B with --speculative-algorithm DSPARK and point --speculative-draft-model-path at the drafter. On Apple silicon, MLX-VLM v0.7.2 or newer accepts the drafter through --draft-model, but DSpark in MLX-VLM currently supports greedy sampling only, so temperature must be set to 0. For llama.cpp, the GGUF drafter pairs with the LFM2.5-VL-3B-GGUF target. Integration work is public in the llama.cpp, SGLang and MLX-VLM pull requests. Acceleration of quantized models is outside the scope of this release.

The release ships under the LFM Open License v1.0, which allows free commercial use only for companies with annual revenue under $10 million. Liquid AI labels it experimental.