general

Jan Releases 30B Multimodal Model for Complex Tasks

Jan releases a 30-billion parameter multimodal AI model designed to handle complex tasks requiring advanced reasoning, visual understanding, and multi-step

Someone found that Jan just dropped a 30B multimodal model that apparently handles long, multi-step tasks better than DeepSeek R1 and Gemini 2.5 Pro.

The model (Jan-v2-VL-max) is built for what they call “long-horizon execution” - basically not losing the plot halfway through complex workflows. It beat both competitors on a benchmark that specifically tests how well models maintain accuracy over extended task chains.

Ways to try it:

For local serving, it works with vLLM 0.12.0 and transformers 4.57.1. The FP8 version comes with production configs already set up, so no need to fiddle with optimization settings.

Released under Apache-2.0 license, which means pretty much unrestricted commercial use. The team hosts the browser version themselves for now while they prep the server repo for release.