Liquid AI发布LFM2.5-VL-3B端侧视觉语言模型
Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device
Yesterday, Liquid AI released LFM2.5-VL-3B. It is a 3.1B-parameter vision-language model built for on-device deployment. The model reads digital screens across mobile, web, and desktop. It grounds objects to coordinates, parses documents and charts, and calls tools from text or image input. Liquid AI reports an average of 69.4 across 28 vision benchmarks. That matches InternVL-3.5-4B and sits 0.7 points behind Qwen3.5-4B, both 4.7B models. The model is non-reasoning, so it answers directly and keeps latency low. It fits in roughly 3 GB of memory and decodes 228 tokens/s on an Apple M5 Max.
昨天，Liquid AI 发布了 LFM2.5-VL-3B。这是一个拥有 31 亿参数的视觉语言模型，专为设备端部署而设计。该模型能够读取移动端、网页端和桌面端的数字屏幕。它可以将对象定位到坐标，解析文档和图表，并从文本或图像输入中调用工具。Liquid AI 报告称，在 28 个视觉基准测试中平均得分为 69.4。这与 InternVL-3.5-4B 相当，比 Qwen3.5-4B 低 0.7 分，两者都是 47 亿参数的模型。该模型不进行推理，因此直接回答并保持低延迟。它大约占用 3 GB 内存，在 Apple M5 Max 上每秒可解码 228 个 token。
Is it deployable?
它可以部署吗？
Yes, the checkpoint ships in four formats: native, GGUF, ONNX, and MLX. Day-one runtimes include llama.cpp, MLX, vLLM, SGLang, and ONNX. It fits in roughly 3 GB of memory.
可以，检查点提供四种格式：原生、GGUF、ONNX 和 MLX。首日运行时包括 llama.cpp、MLX、vLLM、SGLang 和 ONNX。它大约占用 3 GB 内存。
- Which company levels: The LFM Open License v1.0 is Apache-2.0-based with one change: free commercial use ends once a company’s annual revenue reaches $10M USD. So indie developers, startups, and SMBs under that line can ship commercially at no cost. Enterprises above it must negotiate a commercial license with Liquid AI. Research, education, and non-profit use carry no revenue limit.
- Industries: Consumer electronics, automotive, industrial and robotics, financial services, healthcare, and e-commerce. Also QA and RPA vendors that automate GUIs.
- Applications: On-device screen agents, GUI test automation, PDF-to-structured-text with layout labels, invoice and receipt OCR, near-real-time object detection in vehicles, offline translation of menus and road signs, and multi-image comparison.
- 公司级别：LFM 开放许可证 v1.0 基于 Apache-2.0，但有一项变更：一旦公司年收入达到 1000 万美元，免费商业使用即告终止。因此，低于该门槛的独立开发者、初创公司和中型企业可以免费进行商业发布。超过该门槛的企业必须与 Liquid AI 协商商业许可证。研究、教育和非营利用途没有收入限制。
- 行业：消费电子、汽车、工业和机器人、金融服务、医疗保健和电子商务。此外还有自动化 GUI 的 QA 和 RPA 供应商。
- 应用：设备端屏幕代理、GUI 测试自动化、带布局标签的 PDF 到结构化文本、发票和收据 OCR、车辆中的近实时物体检测、菜单和路牌的离线翻译，以及多图像比较。
So, What is new?
那么，有什么新功能？
LFM2.5-VL-3B extends LFM2-VL-3B along four axes.
LFM2.5-VL-3B 在四个维度上扩展了 LFM2-VL-3B。
- Screen and UI understanding: The model averages 80.7 on ScreenSpot-v2 across desktop (78.7), mobile (81.2), and web (82.2). Liquid AI reports Gemma-4-E4B at 51.2 and Qwen3.5-4B at 78.5, with the larger InternVL-3.5-4B ahead at 84.1.
- Function calling: This is new to the VL line. ToolSandbox moves from 26.4 to 59.5. BFCL v4 moves 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 gain driven by scaled synthetic grounding data.
- Multi-image input: BLINK improves from 50.2 to 61.5, and MuirBench from 34.9 to 58.3.
- 屏幕和 UI 理解：该模型在 ScreenSpot-v2 上的平均得分为 80.7，其中桌面端 78.7，移动端 81.2，网页端 82.2。Liquid AI 报告称 Gemma-4-E4B 得分为 51.2，Qwen3.5-4B 得分为 78.5，而更大的 InternVL-3.5-4B 以 84.1 领先。
- 函数调用：这对 VL 系列来说是新增功能。ToolSandbox 从 26.4 提升到 59.5。BFCL v4 从 20.5 提升到 32.5。工具调用以 Pythonic 调用形式在 <|tool_call_start|> 和 <|tool_call_end|> 标记之间发出。
- 定位：RefCOCO 平均 precision@1 从 57.1 提升到 87.9，提升了 30 个百分点，这得益于大规模合成定位数据。
- 多图像输入：BLINK 从 50.2 提升到 61.5，MuirBench 从 34.9 提升到 58.3。
Architecture and training
架构和训练
The language backbone is LFM2.5-2.6B. The vision tower is a SigLIP2 NaFlex shape-optimized 400M encoder. NaFlex handles native resolution by splitting large images into non-overlapping 512×512 patches plus a resized whole-image thumbnail. Context length is 32,768 tokens, and 16 languages are supported.
语言骨干网络是LFM2.5-2.6B。视觉塔是SigLIP2 NaFlex形状优化的400M编码器。NaFlex通过将大图像分割成不重叠的512×512块加上一个调整大小的整图缩略图来处理原生分辨率。上下文长度为32,768个令牌，支持16种语言。
Pre-training used approximately 34T tokens. Vocabulary was doubled to 128K by extending the existing tokenizer in place, which improves non-Latin script coverage. Vision pre-training was scaled 4× in tokens with curated and synthetic caption, OCR, grounding, and instruction-following data.
预训练使用了约34万亿个令牌。词汇量通过就地扩展现有分词器翻倍至128K，这改善了对非拉丁文字符的覆盖。视觉预训练在令牌上扩展了4倍，使用了精选和合成的字幕、OCR、接地和指令跟随数据。
Post-training is SFT with knowledge distillation from a larger teacher and Antidoom training, followed by multi-reward reinforcement learning.
后训练是SFT，使用来自更大教师模型的知识蒸馏和Antidoom训练，随后进行多奖励强化学习。
The model is non-reasoning. It answers directly, which is the design choice behind its latency profile.
该模型是非推理模型。它直接回答，这是其延迟特性背后的设计选择。
Benchmarks
基准测试
Liquid AI evaluated across 28 vision benchmarks using vLLM 0.26.0 in non-reasoning mode. LFM2.5-VL-3B averages 69.4, matching InternVL-3.5-4B (69.4) and landing 0.7 points behind Qwen3.5-4B (70.1). Both comparison models are 4.7B parameters.
Liquid AI在28个视觉基准上使用vLLM 0.26.0以非推理模式进行了评估。LFM2.5-VL-3B平均得分为69.4，与InternVL-3.5-4B（69.4）持平，落后Qwen3.5-4B（70.1）0.7分。两个对比模型均为4.7B参数。
Notable individual results: RealWorldQA 73.1 against InternVL-3.5-4B at 67.7, TextVQA 84.3 against 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 in the prior release.
值得注意的单项结果：RealWorldQA 73.1对比InternVL-3.5-4B的67.7，TextVQA 84.3对比Qwen3.5-4B的81.2，MMStar 63.3，MathVista-mini 68.5，ChartQA 81.3，DocVQA 91.1，OCRBench v1 84.2。CountBenchQA从先前版本的92.2回退到87.3。
On text-only evaluation, IFEval reaches 82.3, up from 72.9. Gemma-4-E4B still leads there at 87.9.
在纯文本评估中，IFEval达到82.3，高于之前的72.9。Gemma-4-E4B仍以87.9领先。
Key Takeaways
关键要点
- LFM2.5-VL-3B hits a 69.4 average across 28 vision benchmarks, matching 4.7B-class models.
- ScreenSpot-v2 jumps to 80.7 and RefCOCO-avg to 87.9, from 57.1 in the prior release.
- Function calling is new to the VL line: ToolSandbox 26.4 → 59.5, BFCL v4 20.5 → 32.5.
- Runs in ~3 GB, decoding 228 tok/s on M5 Max and 20 tok/s on a Galaxy S26 Ultra.
- LFM Open License v1.0 is free commercially only under $10M annual revenue.
- LFM2.5-VL-3B在28个视觉基准上平均得分69.4，与4.7B类模型持平。
- ScreenSpot-v2跃升至80.7，RefCOCO-avg跃升至87.9，而先前版本为57.1。
- 函数调用是VL系列的新增功能：ToolSandbox 26.4 → 59.5，BFCL v4 20.5 → 32.5。
- 运行内存约3 GB，在M5 Max上解码速度为228 tok/s，在Galaxy S26 Ultra上为20 tok/s。
- LFM开放许可证v1.0仅在年收入低于1000万美元时免费商用。
Check out the Technical Details and Model Weights. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看技术细节和模型权重。另外，欢迎在Twitter上关注我们，别忘了加入我们的150k+ ML SubReddit并订阅我们的新闻通讯。等等！你在Telegram上吗？现在你也可以在Telegram上加入我们。
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
需要与我们合作推广你的GitHub仓库或Hugging Face页面或产品发布或网络研讨会等吗？请联系我们
The post Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device appeared first on MarkTechPost.
文章《Liquid AI发布LFM2.5-VL-3B：一款能读取屏幕、定位物体并在设备上调用工具的3B视觉语言模型》首次出现在MarkTechPost上。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力