LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge
liquid.ai
2 threads
Initial benchmarks on my M1 Pro, about 10x faster at describing screenshots than qwen3.5-4b. If the quality is comparably, as the model sheet implies, that’s incredible. But it seems this isn’t a reasoning model?
I worked on this model at Liquid and will be checking in to answer any questions/discussions, both about this vision model and LiquidAI more broadly.
how does this compare to OpenAI nano class models in terms of bench mark scores?
obviously you can't run gpt 5 nano on a mobile device, but if we compare just the performance or things like UI navigation or Text extraction extraction... what is the gap. Is there really a utility gap?