Liquid AI releases vision draft model, with decoding speeds up to 3.13× baseline on M5 Max

Liquid AI released the experimental LFM2.5-VL-3B-DSpark on September 24, adding a draft model with approximately 280 million parameters to an existing vision-language model. It first proposes candidate tokens, which the target model then verifies, aiming to reduce the number of calls required for the target model to decode token by token. Weights and integration entry points for three inference frameworks are available. [Release announcement](https://www.liquid.ai/blog/lfm2-5-vl-dspark)

The draft model extracts hidden states from different layers of the target model. By the time images and text reach these layers, they have been converted into vectors with the same dimensionality, allowing the speculative decoding process from the text version of DSpark to be reused. The model has four layers and was trained with a block length of nine; inference uses eight or nine depending on the hardware. The additional parameters amount to approximately 8.9% of the target model’s parameter count, but this ratio cannot be treated as the memory increase for a complete serving system. [Architecture details](https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspark)

The model card lists six categories of vision tasks. In the COCO benchmark using MLX-VLM on M5 Max, decoding ran at 3.13× the baseline speed, with an overall speedup of 2.59×; the corresponding figures for TextVQA were 2.69× and 1.56×. These Apple benchmarks used FP16, a batch size of one, a temperature of zero, and a block length of eight. The measurements come from the development team; quantized weights and serving under high concurrency require separate validation. [Benchmark tables](https://huggingface.co/LiquidAI/LFM2.5-VL-3B-DSpark)

Integration also required changes to the model interface. An SGLang patch merged on September 22 forwards access to the output head and hidden-state extraction interface required at the top level to the internal language model, preventing missing-member errors during service startup. The output head is exposed through a property to preserve weight key names. Measurements in the PR, using test draft weights, showed that the speedup fell from 1.55× to 1.18× as the batch size increased from eight to 64, indicating that gains may shrink as concurrency rises. [Integration and benchmark records](https://github.com/sgl-project/sglang/pull/40651)

Validation results also require distinguishing algorithm design from implementation. In that SGLang test, the total output token count differed from the baseline, which the author attributed to numerical differences caused by the shapes used during verification. This test alone is insufficient to establish equivalent quality across all responses; deployments should include separate comparisons. [Test limitations](https://github.com/sgl-project/sglang/pull/40651)

In practical applications, image encoding, prefill, and decoding times should be measured separately. Speculative decoding accelerates only the final stage; if image processing accounts for most of the runtime, overall improvements are naturally limited. Engineers should next measure candidate acceptance rates, time to first token, and total latency using charts in Chinese, document question answering, and varying response lengths, while also checking output quality. The model card additionally notes that this path in MLX-VLM currently uses greedy sampling and requires a temperature of zero. [Limitations](https://www.liquid.ai/blog/lfm2-5-vl-dspark), [Runtime settings](https://huggingface.co/LiquidAI/LFM2.5-VL-3B-DSpark)