README.md

Mirage 12B

Mirage represents a novel series of multi-modal models, positioned at a higher tier than the Bumblebee series.

Our preceding 7B model (within the Bumblebee series) had attained a high score on numerous benchmarks. To further enhance its performance, we trained a 12B model. It comprises a ViT vision encoder and a Mistral nemo 12B model.

The results indicate a substantial improvement when compared to many other open-source models.

Our Mirage 12B has surpassed Cambrian-34B and LLava-Next-34B with a single 12B parameter.

The benchmark scores can be viewed on the MMBench official leaderboard.

Roadmap

The currently released version is 2.0 of Mirage. Our 2.1 version is in progress and is likely to be even more potent with multi-lingual OCR support.

If you have any inquiries, feel free to contact us through the issue!