Brave Releases 'Ocelot,' An AI Model For Leo Designed Specifically For One Thing: Summarizing Webpages

In the fast-evolving world of browser-based AI, Brave has introduced a specialized model for its browser.

Calling it the 'Ocelot,' it's designed specifically for web content summarization. Announced in late April 2026, Ocelot is an open-source vision-language model designed from the ground up to handle the unique challenges of summarizing webpages quickly and reliably, all while operating directly within Brave's Leo AI assistant.

When Leo is set to Automatic mode, it selects Ocelot for summarization tasks, or users can choose it manually from the regeneration options below a generated summary.

The model is available on desktop versions of the browser.

Ocelot is a vision-language model based on a LoRA adapter applied to Alibaba's Qwen3-VL-4B-Instruct.

It accepts either plain text from a webpage enclosed in page HTML tags or screenshots of the page. It produces summaries in markdown format and aims to remain neutral and grounded in the source material. The model is trained to respond in the same language as the input content. It is not intended for general conversation, coding, or other tasks outside summarization, and it requires a specific prompt structure to function as designed.

The model includes built-in protections intended to reduce the impact of malicious instructions that may appear in web content.

System-level rules treat page data as untrusted and instruct the model to ignore attempts to override its behavior, change its role, or inject new commands. If the input does not match the expected format, it returns a standard error message asking the user to provide text directly.

Brave provides the model card and weights on Hugging Face under the name Ocelot-1-VL.

The training code, dataset collection tools, and evaluation framework are available on GitHub. The toolkit uses browser automation to gather webpage data in text and image forms, then supports supervised fine-tuning and preference optimization with LoRA. An LLM-as-judge system is included for comparing model outputs during evaluation.

Ocelot operates as part of Brave's broader set of AI options in Leo, alongside other hosted models.

Its design prioritizes efficiency for the specific task of web summarization rather than broad capabilities.

The project is open source on both GitHub and Hugging Face, with the adapter under Apache 2.0 and the training repository under MPL-2.0, allowing external use and modification under the respective licenses. As with other Leo features, summaries are generated locally in the browser context with attention to user privacy.