|

Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device

ByRicardo

Yesterday, Liquid AI launched LFM2.5-VL-3B. It is a 3.1B-parameter vision-language mannequin constructed for on-device deployment. The mannequin reads digital screens throughout cellular, internet, and desktop. It grounds objects to coordinates, parses paperwork and charts, and calls instruments from textual content or picture enter. Liquid AI experiences a mean of 69.4 throughout 28 imaginative and prescient benchmarks. That matches InternVL-3.5-4B and sits 0.7 factors behind Qwen3.5-4B, each 4.7B fashions. The mannequin is non-reasoning, so it solutions straight and retains latency low. It suits in roughly 3 GB of reminiscence and decodes 228 tokens/s on an Apple M5 Max.

Is it deployable?

Yes, the checkpoint ships in 4 codecs: native, GGUF, ONNX, and MLX. Day-one runtimes embrace llama.cpp, MLX, vLLM, SGLang, and ONNX. It suits in roughly 3 GB of reminiscence.

  • Which firm ranges: The LFM Open License v1.0 is Apache-2.0-based with one change: free business use ends as soon as an organization’s annual income reaches $10M USD. So indie builders, startups, and SMBs underneath that line can ship commercially for free of charge. Enterprises above it should negotiate a business license with Liquid AI. Research, schooling, and non-profit use carry no income restrict.
  • Industries: Consumer electronics, automotive, industrial and robotics, financial services, healthcare, and e-commerce. Also QA and RPA distributors that automate GUIs.
  • Applications: On-device display screen brokers, GUI check automation, PDF-to-structured-text with structure labels, bill and receipt OCR, near-real-time object detection in automobiles, offline translation of menus and highway indicators, and multi-image comparability.

So, What is new?

LFM2.5-VL-3B extends LFM2-VL-3B alongside 4 axes.

  • Screen and UI understanding: The mannequin averages 80.7 on ScreenSpot-v2 throughout desktop (78.7), cellular (81.2), and internet (82.2). Liquid AI experiences Gemma-4-E4B at 51.2 and Qwen3.5-4B at 78.5, with the bigger InternVL-3.5-4B forward at 84.1.
  • Function calling: This is new to the VL line. ToolSandbox strikes from 26.4 to 59.5. BFCL v4 strikes from 20.5 to 32.5. Tool calls are emitted as Pythonic calls between <|tool_call_start|> and <|tool_call_end|> tokens.
  • Grounding: RefCOCO-avg precision@1 rises from 57.1 to 87.9, a 30-point acquire pushed by scaled artificial grounding knowledge.
  • Multi-image enter: BLINK improves from 50.2 to 61.5, and MuirBench from 34.9 to 58.3.

Architecture and coaching

The language spine is LFM2.5-2.6B. The imaginative and prescient tower is a SigLIP2 NaFlex shape-optimized 400M encoder. NaFlex handles native decision by splitting massive photographs into non-overlapping 512×512 patches plus a resized whole-image thumbnail. Context size is 32,768 tokens, and 16 languages are supported.

Pre-training used roughly 34T tokens. Vocabulary was doubled to 128K by extending the existing tokenizer in place, which improves non-Latin script protection. Vision pre-training was scaled 4× in tokens with curated and artificial caption, OCR, grounding, and instruction-following knowledge.

Post-training is SFT with information distillation from a bigger instructor and Antidoom training, adopted by multi-reward reinforcement studying.

The mannequin is non-reasoning. It solutions straight, which is the design selection behind its latency profile.

Benchmarks

Liquid AI evaluated throughout 28 imaginative and prescient benchmarks utilizing vLLM 0.26.0 in non-reasoning mode. LFM2.5-VL-3B averages 69.4, matching InternVL-3.5-4B (69.4) and touchdown 0.7 factors behind Qwen3.5-4B (70.1). Both comparability fashions are 4.7B parameters.

Notable particular person outcomes: RealWorldQA 73.1 towards InternVL-3.5-4B at 67.7, TextVQA 84.3 towards Qwen3.5-4B at 81.2, MMStar 63.3, MathVista-mini 68.5, ChartQA 81.3, DocVQA 91.1, and OCRBench v1 84.2. CountBenchQA regressed to 87.3 from 92.2 within the prior launch.

On text-only analysis, IFEval reaches 82.3, up from 72.9. Gemma-4-E4B nonetheless leads there at 87.9.