Liquid AI has introduced LFM2.5-VL-3B, a 3.1B-parameter vision-language model engineered for on-device execution. This model is designed to read digital screens across mobile, web, and desktop environments, enabling capabilities such as object grounding to specific coordinates, parsing documents and charts, and invoking tools based on either text or image inputs. Its architecture prioritizes direct responses over complex reasoning to maintain low latency, fitting within approximately 3 GB of memory and achieving a decoding speed of 228 tokens/s on an Apple M5 Max.
The model's architecture comprises an LFM2.5-2.6B language backbone paired with a SigLIP2 NaFlex shape-optimized 400M vision encoder. The NaFlex component handles native image resolutions by segmenting large images into non-overlapping 512x512 patches, complemented by a resized whole-image thumbnail. LFM2.5-VL-3B supports a context length of 32,768 tokens and 16 languages. Its pre-training involved approximately 34T tokens, with the vocabulary expanded to 128K to enhance non-Latin script coverage. Vision pre-training was scaled fourfold in tokens, utilizing a dataset including curated and synthetic captioning, OCR, grounding, and instruction-following data. Post-training incorporated Supervised Fine-Tuning (SFT) with knowledge distillation from a larger teacher model, Antidoom training, and multi-reward reinforcement learning.
In performance evaluations across 28 vision benchmarks, LFM2.5-VL-3B achieved an average score of 69.4, matching InternVL-3.5-4B and trailing Qwen3.5-4B by 0.7 points; both comparison models are 4.7B parameters. Notable individual results include RealWorldQA at 73.1 (surpassing InternVL-3.5-4B's 67.7) and TextVQA at 84.3 (exceeding Qwen3.5-4B's 81.2). Individual results also included MMStar (63.3), MathVista-mini (68.5), ChartQA (81.3), DocVQA (91.1), and OCRBench v1 (84.2). However, CountBenchQA showed a regression to 87.3 from its prior 92.2. The model is distributed in native, GGUF, ONNX, and MLX formats, with day-one runtime support for llama.cpp, MLX, vLLM, SGLang, and ONNX.
Why this matters
This release is relevant for the deployment of vision-language models to edge devices. The LFM2.5-VL-3B performs screen understanding, object grounding, and tool calling within a 3.1B parameter budget and 3GB memory footprint. Its benchmark scores are competitive with larger 4.7B models (matching InternVL-3.5-4B and closely trailing Qwen3.5-4B), indicating an efficiency-to-performance ratio. The "non-reasoning" design choice, combined with its inference speed on consumer hardware, suggests an approach to latency constraints relevant for real-time on-device applications. The provided format and runtime support may facilitate adoption for developers targeting diverse hardware ecosystems.
- Key Takeaways:
- Parameter Count: 3.1B
- Memory Footprint: ~3 GB
- Inference Speed: 228 tokens/s on Apple M5 Max
- Average Vision Benchmark Score: 69.4 across 28 benchmarks
- Context Length: 32,768 tokens
- Supported Languages: 16
- Checkpoint Formats: native, GGUF, ONNX, MLX
- Runtimes: llama.cpp, MLX, vLLM, SGLang, ONNX