LFM2.5‑VL‑3B Runs a Vision Model in About 3GB of Memory, With Native Support for vLLM, llama.cpp, and MLX

Liquid AI released the LFM2.5‑VL‑3B weights and multiple deployment formats on August 13. Rather than positioning it as a long-chain reasoning model, the company describes it as a low-latency vision-language model designed to run on edge devices. It combines an LFM2.5‑2.6B language backbone with a SigLIP2 NaFlex vision encoder containing roughly 400M parameters, for a total size of about 3.1B parameters. It has a context length of 32,768 tokens and supports 16 languages, including Chinese. Large images are divided into non-overlapping 512×512 tiles, with an additional thumbnail of the full image included to preserve the original aspect ratio and local details.

The main engineering update is that deployment options now span Transformers, vLLM, SGLang, llama.cpp, MLX, ONNX, and GGUF. The model can also output layout OCR results containing region types, normalized coordinates, and content, while representing function calls with special tokens. Liquid AI reports an average ScreenSpot‑v2 score of 80.7, with RefCOCO Precision@1 rising from the previous generation’s 57.1 to 87.9 and ToolSandbox improving from 26.4 to 59.5. These capabilities make it better suited to on-device document extraction, UI grounding, camera-image classification, and lightweight tool-using agents that require visual input.

The performance figures need to be interpreted in light of the test conditions. Liquid AI says the model uses less than 3.3GB of memory and reaches 228, 116, and 20 tokens/s on the M5 Max, Ryzen AI Max+ 395, and Galaxy S26 Ultra, respectively. The roughly 11,000 tokens/s reported for the H100 was measured using vLLM 0.26 and BF16 under a sustained, high-concurrency workload with a single 512×512 image, 1,024 input tokens, and up to 256 output tokens; it is not the speed of a single interactive request. The model card also explicitly advises against using the model for long-context reasoning, visual web design, or question answering over complex blueprints, and notes that the experimental layout format may still change. Developers should next retest Chinese OCR, grounding accuracy after quantization, and peak memory usage across different image sizes. The weights use the custom LFM 1.0 license, whose terms should be reviewed individually before commercial use rather than assuming that “open-weight” means the same as Apache or MIT licensing.