Janus-Pro

A single model for two directions

Janus-Pro is a multimodal model that DeepSeek released on 27 January 2025 as an advanced version of its earlier Janus model^[1]. Its distinguishing feature is that one model does both multimodal understanding — an image goes in, text comes out — and text-to-image generation, where a written prompt produces an image^[2]. Many multimodal systems treat these as two separate jobs; Janus-Pro is built to handle both inside one model.

The design idea inherited from Janus is decoupling. Rather than forcing understanding and generation to share the same visual encoding, Janus splits visual encoding into separate pathways for the two tasks, while keeping a single unified Transformer as the backbone^[1]. In plain terms, the part of the model that has to read an image and describe it, and the part that has to produce an image from a description, no longer have to compromise on a common representation. The Transformer — the sequence model that does the actual processing — stays shared.

Where it came from

The Janus line began before Janus-Pro. The earlier Janus-1.3B appeared in October 2024, and JanusFlow was released on 13 November 2024^[1]. Janus-Pro is the next step in that sequence, and the changes fall into three areas: an optimized training strategy, expanded training data, and a larger model size^[2]. Together these are credited with better multimodal understanding, better instruction-following in text-to-image generation, and more stable generation^[2]. DeepSeek's release notes describe the model as significantly improving multimodal understanding and visual generation^[1], and the arXiv abstract describes significant advancements in both multimodal understanding and text-to-image instruction-following capabilities^[2].

The technical report was posted to arXiv on 29 January 2025, two days after the release^[2].

Sizes and output

Janus-Pro comes in two sizes, Janus-Pro-1B and Janus-Pro-7B, both with a 4,096-token sequence length^[1]. For text-to-image generation the default output is 384x384 pixels, encoded as 576 image tokens per image^[1], and the example configuration uses a patch size of 16^[1]. The arithmetic is easy to check: a 384x384 image cut into 16-pixel patches gives a 24-by-24 grid, which is exactly 576 patches — so the token count follows directly from the resolution and the patch size. Holding an image to 576 tokens is what lets the prompt and the image fit together inside the 4,096-token sequence length.

Licensing and where it sits

The code is MIT-licensed, and use of the model itself is governed by the DeepSeek Model License, which permits commercial use^[1]. That combination matters for anyone who wants to run the model rather than only read about it: permissive code licensing plus a model license that allows commercial deployment lowers the barrier for building on it.

In the wider picture, Janus-Pro is a unified model — not a dedicated image generator and not a dedicated vision-language model, but one system trained in both directions, with the two paths kept separate where separation helps and shared where sharing helps. Its default text-to-image output is 384x384 pixels, so the images it produces by default are small; that is the setting to know before using it.