LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge
16:00 · August 12, 2026 · Hugging Face Blog

Summary
LFM2.5-VL-3B is a 3-billion-parameter vision-language model that pairs a SigLIP2 400M NaFlex vision encoder with the pre-trained backbone of the earlier LFM2.5-2.6B text model. Pre-training consumed roughly 34 trillion tokens and incorporated four times more vision data than prior releases, drawn from curated and synthetic sources covering image captioning, OCR, grounding, and instruction following. The tokenizer vocabulary was extended in place to 128K tokens to improve support for non-Latin scripts without full retraining.
Post-training proceeds in two stages. Supervised fine-tuning incorporates knowledge distillation from a larger teacher model together with Antidoom training, followed by multi-reward reinforcement learning. The resulting model is evaluated on both vision and text benchmarks, including multilingual visual comprehension, document and chart understanding, object grounding, multi-image reasoning, screen navigation, instruction following, and tool use. It leads comparable models on real-world image tasks and matches Gemma-4-E2B and Qwen3.5-2B on tool-calling performance.
Inference measurements show practical deployment characteristics. On-device throughput reaches 228 tokens per second on an M5 Max, 116 tokens per second on a Ryzen AI Max+ 395, and 20 tokens per second on a Galaxy S26 Ultra, with a memory footprint of approximately 3 GB. On an H100, the model sustains roughly 11K output tokens per second under high concurrency, roughly twice the rate of larger 4B-class models and sufficient to produce nearly one billion tokens per day on a single accelerator while maintaining low latency on multi-frame inputs.
Day-one integration is provided for llama.cpp, MLX, vLLM, SGLang, and ONNX, with example code available for multi-image, grounding, OCR, and tool-calling workflows. The model weights are released on Hugging Face.
Why it matters
Directly supplies ML Engineers with architectural details, quantitative benchmarks, latency/memory numbers, and runnable code for deploying efficient VLMs in production or on-device settings.







