deepseek-ai/janus is the repository for Janus-Series, a family of multimodal models that understands text and image inputs and also supports image generation. The repository description states that it unifies multimodal understanding and generation into a single model family. The repository was created on October 18, 2024.
The architecture keeps visual encoding separate and handles both capabilities with a single transformer. According to the README, Janus-Pro scaled up the training strategy, data, and model size, while JanusFlow combined an autoregressive language model with rectified flow. Four models are presented — Janus-1.3B, JanusFlow-1.3B, Janus-Pro-1B, and Janus-Pro-7B — and the README also organizes download and inference examples for each of them. The license is MIT, the primary language is Python, and there are 17761 GitHub stars.
That said, this record has no release version, so the release timing cannot be pinned down, and it provides no benchmark numbers or overall evaluation scores for comparing performance. The only metric left is a momentum score of 2.4 (out of 10), which is low. The separated visual encoding, the single-transformer architecture, and the MIT license can all be confirmed, but the basis for judgment is limited to the README description and the example code. Deciding on real-world use from this record alone is premature, so reviewing the example inference code directly is the more practical step.