# Small vision-language models that run on-device

Which small vision-language model runs on your phone, Mac or in the browser? Compare FastVLM, SmolVLM2, LFM2.5-VL, Gemma 4, Gemma 3n, Qwen3-VL, Moondream, MiniCPM-V and PaliGemma 2 by size, video, browser and MLX support and license.

Canonical URL: https://aiidelist.com/blog/on-device-vlm-comparison

Language: en

Published: 2026-09-24

Updated: 2026-09-24

ON-DEVICE VLM HUB

A vision-language model answers questions about images. The models below are small enough for laptops, phones or WebGPU browsers. Use the table to shortlist, then open a head-to-head comparison with FastVLM.

| Model | Developer | Sizes | Image | Video | Multi-image | Browser | MLX | License | Best for |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| [FastVLM](https://huggingface.co/collections/apple/fastvlm-68ca64a85e3c2a03f32a7098)2025 | Apple | 0.5B · 1.5B · 7B |  |  |  | Yes |  | Apple AMLR | Low time-to-first-token on high-resolution images; browser and Apple Silicon |
| [SmolVLM2](https://huggingface.co/blog/smolvlm2)2025 | Hugging Face | 256M · 500M · 2.2B |  |  |  | Some sizes |  | Apache 2.0 | Tiny models with video understanding; permissive license |
| [LFM2.5-VL](https://huggingface.co/LiquidAI/LFM2.5-VL-3B)2026 | Liquid AI | 450M · 3B |  |  |  | Yes |  | LFM Open License | Grounding, screens, documents and tool calling at the edge |
| [Gemma 4](https://ai.google.dev/gemma/docs/core/model_card_4)2026 | Google | E2B · E4B · 12B · 26B-A4B · 31B |  |  |  | Runtime-dependent |  | Gemma Terms | Broad multimodal (image, video frames, audio on E2B/E4B) and agentic use |
| [Gemma 3n](https://ai.google.dev/gemma/docs/gemma-3n)2025 | Google | E2B · E4B |  |  |  | Runtime-dependent |  | Gemma Terms | Phone-first models with audio input and per-layer embeddings |
| [Qwen3-VL](https://huggingface.co/collections/Qwen/qwen3-vl)2025 | Alibaba | 2B · 4B · 8B (small) |  |  |  | Some sizes |  | Apache 2.0 | Strong OCR, documents and video in small dense sizes |
| [Moondream](https://huggingface.co/vikhyatk/moondream2)2024–2025 | Moondream | 2B · 3 (9B MoE, 2B active) |  |  |  | Some sizes |  | Apache 2.0 / BSL | Pointing, detection and captioning with a tiny footprint |
| [MiniCPM-V](https://huggingface.co/openbmb/MiniCPM-V-4_5)2025 | OpenBMB | 4.5 (8B) |  |  |  | No |  | Apache 2.0 code · MiniCPM weights | High-resolution image and video understanding on laptops |
| [PaliGemma 2](https://huggingface.co/collections/google/paligemma-2)2024 | Google | 3B · 10B · 28B |  |  |  | No |  | Gemma Terms | Fine-tuning base for detection, segmentation and captioning |

“Browser” means a maintained WebGPU/Transformers.js or ONNX build exists for at least one size. “Runtime-dependent” models run through vendor runtimes such as MediaPipe or LiteRT rather than Transformers.js.

Facts checked on 2026-09-17 against official model cards and announcements. Report an outdated row via the site’s GitHub.

## Head-to-head with FastVLM

[FastVLM vs SmolVLM2](/blog/fastvlm-vs-smolvlm2) · [FastVLM vs LFM2.5-VL](/blog/fastvlm-vs-lfm2-5-vl) · [FastVLM vs Gemma 4 / 3n](/blog/fastvlm-vs-gemma) · [FastVLM vs LLaVA-OneVision](/blog/fastvlm-vs-llava-onevision)[Benchmark FastVLM on your device](/blog/benchmark-fastvlm-device)

## How to choose

1. Start from the task: OCR and documents favour higher input resolution; video needs a model trained on frames; grounding needs boxes in the output format.
2. Then the runtime: browser demos need ONNX weights; iPhone apps need MLX or Core ML; Linux servers can use anything.
3. Then the license: Apache 2.0 models are simplest to ship; Apple AMLR and Gemma terms have conditions you must read.
4. Finally measure on your device—published benchmarks use different protocols and cannot be compared across vendors.