AI Trends

Trending story

inclusionAI open-sources Ming-UniVision and MingTok-Vision

1 reports1 story updatesUpdated Nov 17 12:41

Understand this story

AI overview

Ant's inclusionAI has open-sourced Ming-UniVision and MingTok-Vision on GitHub. Ming-UniVision is a 16B-A3B multimodal large model that unifies image understanding, generation, and editing within a single autoregressive next-token prediction framework; MingTok-Vision is a continuous unified vision tokenizer that eliminates discrete quantization. The official announcement states that the shared continuous latent space allows joint training to converge 3.5 times faster, and supports question answering, generation, and semantic editing within multi-turn context without intermediate decoding into pixels. The technical report has been published on arXiv:2510.06590, the model weights are downloadable on Hugging Face and ModelScope, and the repository provides three tiers of running configurations for 44GB, 22GB (int8), and 14GB (int4) of VRAM.

AI-generated from public reports

Latest development · Report

Ant inclusionAI Open-Sources Ming-UniVision and MingTok-Vision Unified Continuous Vision…

Developments

1 confirmed developments

  1. inclusionAI open-sources Ming-UniVision and MingTok-Vision

    Source:English text is being prepared.

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 17

  1. Ant inclusionAI: New GitHub RepositoryFirst-party

    Ant inclusionAI Open-Sources Ming-UniVision and MingTok-Vision Unified Continuous Vision…

    Ant inclusionAI has open-sourced Ming-UniVision and MingTok-Vision on GitHub. The former is a 16B-A3B multimodal large model that unifies image understanding, generation, and editing within a single autoregressive next-token prediction framework; the latter is a continuous unified vision tokenizer that eliminates discrete quantization. The official announcement states that the shared continuous latent space makes joint training converge 3.5 times faster, and supports question answering, generation, and semantic editing in multi-turn contexts without intermediate decoding into pixels. The technical report has been published on arXiv:2510.06590, the model weights are downloadable on Hugging Face and ModelScope, and the repository provides three tiers of memory usage: 44GB, 22GB (int8), and 14GB (int4).

Related stories

  1. Related storiesMiniMax releases the H3-Image image model and opens the weights
  2. Related storiesHi3D open-source image-to-3D model Hi3D v1.0
  3. Related storiesAnt Group open-sources ARGenSeg inference and training code and checkpoints
  4. Related storiesAnt Group's GuangJian team releases visual forensics model DefakerOne2