Liquid AI LFM2.5-VL-3B: 3.1B On-Device Vision-Language Model
In this article
Liquid AI has released LFM2.5-VL-3B, a 3.1B-parameter vision-language model designed for on-device deployment across mobile, web, and desktop environments. The model reads digital screens, grounds objects to normalized bounding-box coordinates, parses documents and charts, and — for the first time in Liquid AI's VL line — emits Pythonic tool calls triggered from either text or image input. It fits in approximately 3 GB of memory and decodes at 228 tokens per second on an Apple M5 Max, making it a practical option for builders evaluating small language models for production edge deployments.
The headline number is a 69.4 average across 28 vision benchmarks, evaluated using vLLM 0.26.0 in non-reasoning mode. That ties InternVL-3.5-4B at 69.4 and trails Qwen3.5-4B by 0.7 points — both are 4.7B-parameter models. The non-reasoning design answers directly without generating a reasoning trace, compressing time-to-first-token. On a single NVIDIA H100 SXM5 with a 5-frame 256×256 clip, TTFT measures approximately 34 ms; Liquid AI places Gemma-series models near 200 ms on the same task.
Architecture
The language backbone is LFM2.5-2.6B. The vision encoder is a SigLIP2 NaFlex shape-optimized model at 400M parameters. NaFlex handles variable-resolution input by splitting images into non-overlapping 512×512 patches and appending a single resized whole-image thumbnail, preserving spatial fidelity for large screenshots without a fixed-resolution bottleneck. Context length is 32,768 tokens spanning text and vision tokens combined.
Pre-training consumed approximately 34T tokens. The tokenizer vocabulary was doubled to 128K by extending the existing tokenizer in place, improving non-Latin script coverage across 16 supported languages. Vision pre-training was scaled 4× in tokens relative to the prior release, using curated and synthetic data across captioning, OCR, grounding, and instruction-following. Post-training combined supervised fine-tuning with knowledge distillation from a larger teacher model and Antidoom training, followed by multi-reward reinforcement learning.
Benchmark Results
The sharpest gains relative to LFM2-VL-3B appear in grounding and UI understanding. RefCOCO-avg precision@1 moved from 57.1 to 87.9, a 30-point jump attributed to scaled synthetic grounding data; coordinates are output as normalized integers in the 0–1000 range. On ScreenSpot-v2, the model averages 80.7: 78.7 on desktop, 81.2 on mobile, and 82.2 on web.
| Benchmark | LFM2-VL-3B (prior) | LFM2.5-VL-3B (3.1B) | InternVL-3.5-4B (4.7B) | Qwen3.5-4B (4.7B) | Gemma-4-E4B |
|---|---|---|---|---|---|
| 28-bench avg | — | 69.4 | 69.4 | 70.1 | — |
| ScreenSpot-v2 | — | 80.7 | 84.1 | 78.5 | 51.2 |
| RefCOCO-avg P@1 | 57.1 | 87.9 | 88.8 | — | — |
| ToolSandbox | 26.4 | 59.5 | — | — | — |
| BFCL v4 | 20.5 | 32.5 | — | — | — |
| IFEval (text-only) | 72.9 | 82.3 | — | — | 87.9 |
| DocVQA | — | 91.1 | — | — | — |
| TextVQA | — | 84.3 | — | 81.2 | — |
| RealWorldQA | — | 73.1 | 67.7 | — | — |
| BLINK (multi-image) | 50.2 | 61.5 | — | — | — |
| MuirBench (multi-image) | 34.9 | 58.3 | — | — | — |
| CountBenchQA | 92.2 | 87.3 | — | — | — |
CountBenchQA is the one notable regression: 87.3 versus 92.2 in the prior release. InternVL-3.5-4B leads on both ScreenSpot-v2 (84.1) and RefCOCO-avg (88.8), so engineers who can absorb a larger footprint retain a performance ceiling above LFM2.5-VL-3B on those tasks.
Deployment and Licensing
The checkpoint ships simultaneously in four formats — native, GGUF, ONNX, and MLX — with day-one runtime support across llama.cpp, MLX, vLLM, SGLang, and ONNX. The function-calling interface emits Pythonic calls between <|tool_call_start|> and <|tool_call_end|> tokens, relevant to developers building agentic pipelines where tool invocation reliability matters.
Licensing follows LFM Open License v1.0, an Apache-2.0 derivative with a revenue threshold: free commercial use applies only to organisations generating under $10M USD annually. Startups, indie developers, and SMBs below that line can ship without a fee; enterprises above it must negotiate a separate commercial license. Research, education, and non-profit use carry no revenue ceiling.
LFM2.5-VL-3B occupies the gap where on-device agentic behaviour intersects with constrained memory budgets — territory that 7B-class models cannot fill and pure text models cannot address. ToolSandbox's jump from 26.4 to 59.5 in a single release cycle, combined with multi-image gains on BLINK and MuirBench, indicates the architecture has room still to grow.