A 280M-parameter model speeds up LFM2.5-VL without changing its outputs

Liquid AI is extending its DSpark speculative decoding approach to vision-language models with an experimental drafter for LFM2.5-VL-3B. The method reaches up to 3.13x faster decoding on Apple Silicon and 2.66x on H100 while preserving the target model's output distribution.

Predict several tokens before verifying them

Making a vision-language model respond faster does not necessarily require modifying the model itself. With LFM2.5-VL-DSpark, Liquid AI is extending to multimodal inference a speculative decoding technique already used with its text models.

A much smaller auxiliary model predicts several tokens ahead using hidden states from LFM2.5-VL-3B. The target model then verifies that block in a single pass instead of generating every token sequentially.

From the drafter's perspective, the distinction between text and images has largely disappeared at this stage of computation. Text tokens and image representations exist as tensors within the model's hidden layers, allowing Liquid AI to apply the same underlying approach used with its text models. 279.5 million additional parameters

The selected drafter contains four attention-only layers and 279.5 million parameters. Adding it increases the deployed system's parameter count by 8.9%. Embeddings and the LM head are shared with the target model rather than duplicated.

Liquid AI trained several variants before selecting a block size of 9 during training and either 8 or 9 at inference depending on the hardware. All drafter experiments and ablations were trained on AMD hardware using Liquid AI's training framework.

The checkpoint is available on Hugging Face, alongside a GGUF version, with inference support across llama.cpp, MLX-VLM, and SGLang. Up to 3.13x faster on Apple Silicon

Liquid AI evaluated DSpark across six categories of vision tasks, including general VQA, Text VQA, image captioning, Chart VQA, complex reasoning, and multi-turn conversation.

On an M5 Max MacBook Pro using MLX, decoding is between 2.30x and 3.13x faster depending on the task. End-to-end improvements range from 1.56x to 2.62x.

Using llama.cpp on an M3 Ultra, decoding improves by 1.57x to 2.14x, while end-to-end gains range from 1.30x to 1.77x.

On a single H100 80GB running SGLang, Liquid AI reports 2.04x to 2.66x faster decoding and 1.64x to 2.27x end-to-end improvements.

These measurements were collected at batch size 1 and temperature 0, using FP16 on Apple Silicon and BF16 on H100. Liquid AI collected the performance data through its Pipette benchmarking infrastructure. Faster inference without changing the output distribution

Speculative decoding does not simply accept whatever the smaller model predicts. Every proposed token remains subject to verification by LFM2.5-VL-3B. When a proposal does not match its distribution, the target model rejects it and samples a corrected replacement.

Under matched sampling settings, this procedure preserves the target model's distribution. Liquid AI therefore describes the process as lossless: the drafter changes how generation is computed, not the model the final output comes from.

Efficiency still depends on acceptance rate. The more accurately the drafter anticipates upcoming tokens, the more tokens can be validated during each target-model pass. Higher temperatures increase disagreement between the drafter and target, reducing throughput gains. Vision processing remains outside the acceleration

DSpark does not accelerate every stage of vision-language inference. Before text generation begins, an image still has to pass through the vision encoder, followed by a prefill stage processing the resulting visual tokens.

The drafter only accelerates decoding. A workload dominated by initial image processing will therefore benefit less than one that generates a long response after the image has been encoded.

Liquid AI's measurements show the difference directly. Decoding reaches up to 3.13x the baseline speed on M5 Max, while the maximum end-to-end gain falls to 2.62x. On M3 Ultra, the corresponding maximums are 2.14x and 1.77x.

The practical speedup therefore depends less on the presence of an image itself than on how much of the total workload is spent generating tokens after that image has been processed.