Liquid AI says its vision model can run up to 3.13 times faster on-device

Liquid AI has released an experimental draft model that it says can accelerate its LFM2.5-VL-3B vision-language model by up to 3.13 times on an Apple M5 Max. The important bit is not merely a larger benchmark number: the release brings speculative decoding to a model designed to understand images, documents and screens, with support for llama.cpp, MLX-VLM and SGLang from day one.

Watch Desk analysis

What happened

Liquid AI says the LFM2.5-VL-DSpark drafter adds about 280 million parameters, or 8.9 per cent, to the 3B target model. It proposes blocks of tokens which the target model then verifies, preserving the target’s output under greedy decoding while attempting to reduce the time spent generating responses.

In Liquid AI’s published results, decoding on an M5 Max with MLX was between 2.30 and 3.13 times faster across six vision tasks. End-to-end latency improved by 1.56 to 2.62 times. Tests using llama.cpp on an M3 Ultra showed decoding gains of 1.57 to 2.14 times and end-to-end gains of 1.30 to 1.77 times. On an NVIDIA H100, Liquid AI reports decoding improvements of up to 2.66 times and end-to-end gains of up to 2.27 times.

The model is available in Safetensors and GGUF formats, with launch instructions for llama.cpp, MLX-VLM and SGLang. That makes this more useful than a laboratory flourish: developers can try the drafter in several established local and GPU-serving stacks.

Why it matters

Vision-language models have an awkward latency problem. They must first process an image, then work through visual tokens before generating an answer. Speculative decoding speeds up token generation, but it does not accelerate image encoding or prompt prefill. Liquid AI acknowledges that those untouched stages can cap the real-world benefit, particularly when the input is visually complex.

For developers running AI on laptops, workstations or other edge hardware, the practical promise is a snappier model without changing the target model’s verified output. The catch is that the headline decoding gain will not necessarily become the same improvement in the time a person waits for an answer. Physics remains an annoyingly persistent product manager.

Our read

This is a credible and technically interesting move towards faster on-device vision AI. The strongest part of the release is its breadth of integrations and its clear explanation of where speculative decoding helps, and where it cannot.

Treat the figures as Liquid AI’s own measurements, not a universal speed guarantee. Hardware, task mix, prompt length and serving configuration will decide whether the improvement feels transformational or merely pleasant. If the approach holds up outside the company’s test setup, small vision models become more attractive for privacy-sensitive and bandwidth-limited applications.

What to watch

• Independent tests on ordinary laptops, phones and low-power edge devices.
• Whether end-to-end gains remain strong on long documents and image-heavy prompts.
• Support arriving in stable builds of llama.cpp, MLX-VLM and SGLang.
• Whether other model makers adopt the same drafter approach for vision-language systems.

Discussion spark: Should developers prioritise faster local vision models even when the biggest gains are in decoding rather than total response time?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.