New open AI models

Thursday, October 15

4 new models on Hugging Face that day.

• Chat and writingtencent

  Hy-Memory-30B-A3B

  A Tencent model that writes long-term memory for AI agents by picking what to keep from their conversations.

  • Pulls out the facts worth remembering from conversations and words them clearly
  • 30B parameters in total, only about 3B active per token
  • Built to plug into memory tools like Mem0 and MemOS, not for general chat

  Minimum 24GB graphics card e.g. GeForce RTX 5090or a computer with 32GB memory and no graphics card (slower)

  Free for commercial use

  Download ↗

• Search and matchingnvidia

  Nemotron-3.5-Rerank-8B-BF16

  An 8B NVIDIA reranker that sorts search results by how well they match a question, for text or page images.

  • Scores how relevant each document is to a text query, including charts and tables
  • Evaluated on 34 languages for text and 18 languages for page images
  • It only re-orders results and does not write answers; a search system must feed it

  Minimum 8GB graphics card, or 16GB computer memory (slower) e.g. GeForce RTX 5050

  Free for commercial useSupports Arabic

  Download ↗

• Speaks textSupraLabs

  SupraTTS-1.0-Preview

  A very small model that turns English text into speech, with about one million parameters.

  • On a plain CPU it runs about 90 times faster than real time, the maker says
  • Needs only Python and three pip packages to run
  • English only, a preview release, and the quality scores are the maker's own

  Minimum A recent phone or any computer

  Free for commercial use

  Download ↗

• 3D shapesAlayaLab

  MeshAxiom-Base

  A model that turns one picture of an object into a 3D mesh that can be saved as OBJ or GLB.

  • Input is a 512×512 RGB image on a black background
  • Exports OBJ, JSON and GLB; the requested face count is only a guide
  • Weights only: it needs the separate MeshAxiom code package to run

  Minimum 24GB graphics card e.g. GeForce RTX 5090

  Check the licence before use

  Download ↗

The minimum is our estimate from the model’s size: a 4-bit copy for text models, 8-bit for image, video and sound. Licences and sizes come from each maker’s page; read the licence before using a model in your work.