LFM2.5-VL-3B aims to make capable vision-language inference practical on edge devices: it trades extensive internal chain-of-thought reasoning for fast first-token response, compact memory use (~3 GB), and multi-feature support (OCR with layout, grounding, bounding boxes, and tool/function calls). The design choice favors single-turn, high-throughput visual tasks where latency and deployability matter more than heavy multimodal reasoning.
Key Capabilities
- Compact, device-aware architecture: combines a 2.6B‑scale language backbone with a SigLIP2 NaFlex 400M vision encoder so you can run vision-text inference on laptops and phones (examples: ~228 tokens/s on Apple M5 Max, ~116 tokens/s on AMD Ryzen AI Max+ 395, ~20 tokens/s on Galaxy S26 Ultra) — so what: you can deploy private multimodal features without large GPU clusters.
- Structured OCR and layout output: returns full-page OCR with labeled regions and normalized bounding boxes in a documented layout annotation format — so what: easier pipeline integration for PDF→searchable text, document parsing, or downstream extraction tasks.
- Strong grounding and tool use: improved synthetic grounding data and tool/function calling support (Pythonic tool-call tokens + tool role round-trips) that raise RefCOCO precision and sandbox/toolbench metrics — so what: better accuracy for object-level queries and automated tool workflows that require structured calls and results.
- Long-context and multilingual tuning: 32,768 token context and a 128k vocabulary with multilingual coverage — so what: can handle larger visual-document inputs and many scripts for translation/understanding tasks on-device.
Who It's For and Trade-offs
Great fit if you need near-realtime, single-turn visual understanding on constrained hardware (mobile, embedded, or CPU servers), e.g., on-device OCR with layout, UI-screen understanding, or fast object grounding and tool-driven pipelines. Look elsewhere if your workload requires deep chain-of-thought multimodal reasoning, very long multi-turn visual dialogs, or highest-end benchmark-leading multimodal reasoning — the model is intentionally optimized for responsiveness and footprint rather than extensive internal reasoning.
Where It Fits
Positioned between heavy cloud multimodal models and tiny mobile encoders: it outperforms prior LFM2-VL-3B in grounding, screen understanding and tool use while remaining compact and exportable to formats like GGUF, ONNX and MLX for diverse deployment targets. Use it when latency, memory, and deployability are first-order constraints.
