Reading · Developers Digest

This site does not allow itself to be embedded. You can still read it on the original site — the toolbar below keeps your place in the directory.

Liquid AI released LFM2.5-VL-3B on August 12, 2026: a 3.1B open-weights vision-language model that averages 80.7 on ScreenSpot-v2, doubles ToolSandbox to 59.5, and decodes at 228 tokens/s on an M5 Max in about 3 GB of memory. Here is what shipped, the benchmark caveats, and how to run it.

On August 12, 2026, Liquid AI released LFM2.5-VL-3B, its most capable open-weights vision-language model. The 3.1B-parameter model reads digital screens (80.7 average on ScreenSpot-v2, 29.5 points above the 8B Gemma-4-E4B), grounds objects at 87.9 RefCOCO precision@1, and calls tools at 59.5 on ToolSandbox. It is a non-reasoning model that answers directly, which is why it decodes at 228 tokens/s on an M5 Max and 116 tokens/s on a Ryzen AI Max+ 395 in about 3 GB of memory. That combination - screen understanding, grounding, and function calling under 3.3 GB - is the release worth reading about. ## Official Sources | Resource | Description | |----------|-------------| | [Liquid AI blog: LFM2.5-VL-3B release post](https://www.liquid.ai/blog/lfm2-5-vl-3b) | Training pipeline, full benchmark tables, inference measurements | | [Hugging Face: LiquidAI/LFM2.5-VL-3B model card](https://huggingface.co/LiquidAI/LFM2.5-VL-3B) | Architecture details, license, layout annotation format, demos | | [Hugging Face blog: LFM2.5-VL-3B post](https://huggingface.co/blog/LiquidAI/lfm2-5-vl-3b) | Runnable transformers example, WebGPU demo, ecosystem support | | [Liquid AI docs: vision capabilities](https://docs.liquid.ai/lfm/key-concepts/vision-capabilities) | Multi-image, grounding, OCR, and tool-calling examples | | [WebGPU browser demo](https://huggingface.co/spaces/LiquidAI/LFM2.5-VL-3B-WebGPU) | Vision chat with grounding boxes and tool calls, no setup | ## What Shipped LFM2.5-VL-3B pairs the same pre-trained backbone as the [LFM2.5-2.6B text model](/blog/lfm2-5-2-6b-on-device-agentic-model) with a SigLIP2 400M NaFlex vision encoder. Pre-training used about 34T tokens with 4x more vision data than the preceding LFM2-VL-3B: curated and synthetic image-caption, OCR, grounding, and instruction-following sets. The tokenizer vocabulary doubled to 128K in place to support non-Latin scripts. Post-training runs supervised fine-tuning with knowledge distillation from a larger teacher plus Antidoom training, then multi-reward reinforcement learning. Four capabilities improved versus the previous VL release: 1. **Screen/UI understanding.** 80.7 average on ScreenSpot-v2 across desktop, mobile, and web, versus 51.2 for the 8B gemma-4-E4B-it and 78.5 for the 4.7B Qwen3.5-4B. 2. **Function calling, new to the VL line.** ToolSandbox more than doubles from 26.4 to 59.5; BFCL V4 climbs from 20.5 to 32.5, putting it on par with Gemma-4-E2B and ahead of Qwen3.5-2B. 3. **Grounding.** RefCOCO precision@1 jumps from 57.1 to 87.9, a 30-point gain on synthetic grounding data. 4. **Multi-image input.** BLINK rises from 50.2 to 61.5 and MuirBench from 34.9 to 58.3. ## The Benchmarks Vendor-published numbers, normalized 0-100, evaluated with vLLM 0.26.0 in non-reasoning mode. Selected rows: | Benchmark | LFM2.5-VL-3B (3.1B) | LFM2-VL-3B (3.1B) | gemma-4-E4B-it (8B) | Qwen3.5-4B (4.7B) | |-----------|----------------------|--------------------|---------------------|--------------------| | MMStar | 63.3 | 57.7 | 52.9 | 59.3 | | RealWorldQA | 73.1 | 71.1 | 64.3 | 67.1 | | DocVQA (val) | 91.1 | 89.8 | 87.4 | 94.8 | | TextVQA (val) | 84.3 | 83.0 | 69.0 | 81.2 | | ChartQA (test) | 81.3 | 80.4 | 42.1 | 84.2 | | RefCOCO-avg | 87.9 | 57.1 | 72.1 | 86.6 | | ScreenSpot-v2 Desktop | 78.7 | 6.0 | 45.8 | 76.3 | | ScreenSpot-v2 Mobile | 81.2 | 7.6 | 60.3 | 81.4 | | ScreenSpot-v2 Web | 82.2 | 2.5 | 47.6 | 77.8 | | BLINK | 61.5 | 50.2 | 52.2 | 58.7 | | IFEval | 82.3 | 72.9 | 87.9 | 86.2 | | ToolSandbox | 59.5 | 26.4 | 61.6 | 65.0 | | BFCL V4 | 32.5 | 20.5 | 40.0 | 53.6 | | Average (all 28) | 69.4 | 57.2 | 59.7 | 70.1 | The pattern matches the training story: LFM2.5-VL-3B leads its size class on real-world image tasks and is strongest where the release claims focus - screens, grounding, and tool use - while the 8B Gemma and 4.7B Qwen stay ahead on instruction following and the heaviest function-calling suites. These are vendor measurements with vendor generation parameters; treat the 0.7-point gap to Qwen3.5-4B as a claim, not an independent verdict. ## Inference Speed Day-one support covers llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX. On-device, the model decodes 228 tokens/s on an M5 Max and 116 tokens/s on a Ryzen AI Max+ 395 within about 3.3 GB of memory, and reaches 20 tokens/s on a Galaxy S26 Ultra. On a single H100 the answer-direct design shows up in latency: about 34 ms to first token on a 5-frame video clip where the Gemma models take around 200 ms, plus roughly 11K output tokens/s at high concurrency, about 2x the 4B-class models and close to 1B output tokens per day. ## How to Run It The model id is `LiquidAI/LFM2.5-VL-3B`, and a working transformers path (requires `transformers>=5.10.1`): ```python from transformers import AutoModelForImageTextToText, AutoProcessor model_id = "LiquidAI/LFM2.5-VL-3B" processor = AutoProcessor.from_pretrained(model_id) model = AutoModelForImageTextToText.from_pretrained( model_id, device_map="auto", dtype="bfloat16", ) ``` The [WebGPU space](https://huggingface.co/spaces/LiquidAI/LFM2.5-VL-3B-WebGPU) is the fastest way to judge the model: upload images, watch it return bounding boxes and tool calls. Fine-tuning notebooks live in the [Liquid4All cookbook](https://github.com/Liquid4All/cookbook/tree/main/finetuning/notebooks), under the LFM Open License v1.0. OpenCode does not list LFM2.5-VL-3B in its model registry, so there is no `opencode run --model` one-liner. As with the sibling [LFM2.5-2.6B](/blog/lfm2-5-2-6b-on-device-agentic-model), the vendor path is to serve it with vLLM or llama.cpp and wire it in as a custom provider. ## Why It Matters The previous [on-device agent model post](/blog/lfm2-5-2-6b-on-device-agentic-model) argued that local inference changes what you run because marginal cost drops to zero. This release adds the missing modality. Screen understanding plus grounding plus function calling in a 3B package is the ingredient list for GUI agents that run on the machine they automate: the model can see a desktop, locate the button, and invoke the tool without sending a screenshot to a cloud API. For document workflows the numbers are already production-usable, with DocVQA at 91.1 and layout-aware OCR covered in the model card. Two caveats keep it honest. First, the vision suite average (69.4) trails the 4.7B Qwen3.5 by 0.7 points, so "outperforms larger models" is selective: it is genuinely ahead of the 8B Gemma, but Qwen3.5-4B remains the stronger generalist. Second, function calling still favors bigger models (BFCL V4: 32.5 versus 53.6), so tool-heavy agent loops on server GPUs should stay larger. The sweet spot is private, on-device, high-volume vision work: screenshot analysis, UI automation, OCR-heavy pipelines, and phone or laptop deployments where latency and data residency matter more than benchmark-topping. Compare that against [Claude's computer use](/blog/claude-computer-use), which trades away local execution for a much stronger reasoning ceiling. ## FAQ ### Is LFM2.5-VL-3B open weights? Yes. Weights are on Hugging Face under the LFM Open License v1.0: commercial use below $10M annual revenue, free for non-profits and research. Above the threshold you need a commercial agreement. ### What hardware does LFM2.5-VL-3B need? About 3.3 GB of memory on device. Liquid AI measured 228 tokens/s on an M5 Max, 116 tokens/s on a Ryzen AI Max+ 395, and 20 tokens/s on a Galaxy S26 Ultra. On GPU it serves with vLLM or SGLang at roughly 11K output tokens/s per H100. ### Can LFM2.5-VL-3B call tools? Yes, and it is the headline addition to the VL line: ToolSandbox at 59.5 (up from 26.4) and BFCL V4 at 32.5. It calls tools on text-only and vision-plus-text inputs and can return grounding bounding boxes alongside. ### How does it compare to Qwen3.5-4B? Liquid AI measures 69.4 average versus 70.1 across the suite. LFM2.5-VL-3B wins on screens and grounding; Qwen3.5-4B stays ahead on function calling and instruction following. ## Continue Reading - [LFM2.5-2.6B: Liquid AI's On-Device Agent Model](/blog/lfm2-5-2-6b-on-device-agentic-model) - the text-only sibling, same family and license - [The Best Local Coding LLMs in 2026](/blog/best-local-coding-llms-2026) - where on-device models stand on real workloads - [Claude Computer Use](/blog/claude-computer-use) - the cloud-side approach to UI automation - [GLM 5.2 on a Slow Computer: Local Inference](/blog/colibri-glm-52-slow-computer-local-inference) - what local inference costs in practice - [What Is an AI Coding Agent in 2026](/blog/what-is-an-ai-coding-agent-2026) - how harnesses, tools, and models fit together - [TurboFieldfare: Running Gemma 4 26B in 2 GB of RAM on Any M-Series Mac](/blog/turbo-fieldfare-gemma-4-26b-2gb-ram-mac) ## Sources - [Liquid AI: LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge](https://www.liquid.ai/blog/lfm2-5-vl-3b) - fetched August 12, 2026 - [Hugging Face blog: LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge](https://huggingface.co/blog/LiquidAI/lfm2-5-vl-3b) - fetched August 12, 2026 - [Hugging Face: LiquidAI/LFM2.5-VL-3B model card](https://huggingface.co/LiquidAI/LFM2.5-VL-3B) - fetched August 12, 2026 - [Liquid AI docs: LFM vision capabilities](https://docs.liquid.ai/lfm/key-concepts/vision-capabilities) - fetched August 12, 2026

LFM2.5-VL-3B: Liquid AI's 3B Vision Model Reads Screens, Grounds Objects, and Calls Tools on a Laptop

Comments