Liquid AI releases LFM2.5-VL-DSpark to speed up vision models by 3x

Liquid AI has released an experimental draft model called LFM2.5-VL-DSpark. This tool is designed to make their vision-language model, LFM2.5-VL-3B, run significantly faster. The release note details how this new component integrates with existing hardware to improve performance.

The core idea is speculative decoding. Instead of generating one word at a time, the model predicts a whole block of words in advance. A smaller "drafter" model guesses these words, and the main model checks them. If the guesses are correct, the process skips several steps, saving time.

This approach requires a small memory increase. The drafter adds 280 million parameters to the system. That is an 8.9% increase on top of the original 3 billion parameter model. In exchange, users get a major speed boost.

On Apple M5 Max chips, decoding runs up to 3.13 times faster. On high-end H100 GPUs, the speedup reaches 2.66 times. End-to-end latency, which includes the time to process the image, also improves by up to 2.62 times on local devices.

The technical trick involves shared representations. Image patches and text tokens are projected into the same format before reaching the drafter. This allows the same inference algorithm used for text to work seamlessly with images.

Developers can use this model immediately. Liquid AI provided day-one support for three popular frameworks: llama.cpp, MLX-VLM, and SGLang. These tools allow the model to run on both local computers and cloud servers.

However, the speedup has limits. Speculative decoding only accelerates the text generation part of the process. It does not speed up the initial analysis of the image. On devices with less power, this image processing step takes a larger share of the total time.

Because of this, the overall speed gain is smaller than the raw decoding speed. The source notes that this is a standard limitation known as Amdahl's law. The unaccelerated part of the task caps the total improvement.

The model is available in Safetensors and GGUF formats on Hugging Face. It is open-weight, meaning users can download, fine-tune, and deploy it without restrictions. This fits Liquid AI's goal of providing AI that runs anywhere.

The release is part of a broader effort to make vision models practical for everyday use. By reducing latency, these models become more responsive in real-time applications. This is crucial for interactive tools that need quick answers.

The source does not provide details on how this model performs on older hardware. It also does not specify the exact accuracy metrics for the draft model itself. The focus remains on speed and integration rather than new capabilities.

Does a 3x speedup change how you plan to use vision models on your own devices?