A lightweight, open vision-language model built to understand and summarize home security camera events.
``` ollama run llmvision/glimpse-v1 ``` > Install [Ollama](https://ollama.com), then paste the command above. See the [project site](https://llmvision.org/glimpse/) for documentation. ## Model summary | | | |---|---| | **Developer** | LLM Vision | | **Base model** | `google/gemma-3-4b-pt` | | **Architecture** | Gemma 3 (vision-language) | | **Parameters** | ~4B | | **Modality** | Image + text → text | | **Training samples** | 5,000+ real-world home security camera events | | **Reported gain** | 1.9× accuracy improvement over the base model on the target task | | **License** | [Gemma Terms of Use](https://ai.google.dev/gemma/terms) | ## Intended use Glimpse-v1 is purpose-built for **summarizing and describing footage from home security cameras** — for example, generating short natural-language descriptions of motion events, deliveries, visitors, pets, or vehicles, locally on consumer hardware. ### Designed for - Local, privacy-preserving smart-home automations using LLM Vision - Event summaries for camera notifications - Edge devices and machines with limited VRAM/RAM ### Not designed for - General-purpose visual question answering or document understanding - Person identification, biometric recognition, or surveillance of identifiable individuals - Safety-critical decisions (medical, legal, security response) without human review - Use cases prohibited by the [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy) ## Languages English only for now. ## Why a small model? Glimpse-v1 is a **compact 4B-parameter** model deliberately sized to run on hardware with limited memory and compute. The goal is **private, local AI for the home**: your camera footage never has to leave your network, and you avoid recurring API costs. ## Performance Glimpse-v1 reports a **1.9× accuracy improvement** over the base Gemma 3 4B model on home-security event summarization. See the [project site](https://llmvision.org/glimpse/) for the latest benchmarks. ## Training - **Base:** Gemma 3 4B - **Data:** ~5,000 curated real-world home security camera events spanning diverse scenes, lighting conditions, and event types - **Objective:** Supervised fine-tuning for concise, factual event descriptions > Files in this repository have been **modified from the original Gemma 3 release** as part of this fine-tune. ## How to use ### Ollama (recommended) ``` ollama run llmvision/glimpse-v1 ``` ### Transformers ``` from transformers import AutoProcessor, AutoModelForImageTextToText import torch model_id = "llmvision/glimpse-v1" processor = AutoProcessor.from_pretrained(model_id) model = AutoModelForImageTextToText.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto" ) messages = [ {"role": "user", "content": [ {"type": "image", "url": "path/to/frame.jpg"}, {"type": "text", "text": "Summarize this camera event in one sentence."}, ]}, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) out = model.generate(inputs, max_new_tokens=128) print(processor.decode(out[0], skip_special_tokens=True)) ``` ## Limitations and risks - **Domain-specific.** Outside of home-security framing, quality drops noticeably. - **Hallucination.** Like all VLMs, it can invent details (people, objects, actions) not present in the image. Treat outputs as suggestions, not ground truth. - **Bias.** Training data reflects the distribution of available home camera footage and may underperform on under-represented scenes, lighting, or demographics. - **Privacy.** Although the model runs locally, **you** are responsible for handling footage of identifiable people in line with local laws (e.g. GDPR). - **Not a security system.** Do not use Glimpse-v1 as the sole signal for emergency response. ## License This model is a **Gemma Model Derivative** and is distributed under the [**Gemma Terms of Use**](https://ai.google.dev/gemma/terms). Use, reproduction, modification, and redistribution are subject to those terms, including the [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy). By downloading or using Glimpse-v1 you agree to the Gemma Terms of Use. If you redistribute Glimpse-v1 or any derivative of it, you must: 1. Pass these terms through to your recipients as an enforceable provision. 2. Provide recipients a copy of the Gemma Terms of Use. 3. Mark any modified files with prominent notices that they have been modified. 4. Include a `NOTICE` file containing: > Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. ## Citation ``` @misc{glimpse_v1, title = {Glimpse-v1: A compact vision-language model for home security event understanding}, author = {Valentin Fröhlich}, year = {2026}, url = {https://llmvision.org/glimpse/} } ``` ## Acknowledgements Built on [Google Gemma 3](https://ai.google.dev/gemma). Distributed via [Ollama](https://ollama.com) and Hugging Face.