CPU3DAI tools for 3D, video and audio

Tencent releases Hy-Memory-30B-A3B for external memory of AI agents

Tencent Hunyuan has published the Hy-Memory-30B-A3B model on HuggingFace for the text-generation task. It is a compact model with roughly 3 billion active parameters, trained to write data into external memory systems — Mem0, MemOS and Hy-Memory. It decides what to save, how to represent information and how to update the memory bank with new data.

What it means

The full model contains 30.1 billion parameters; the weights in the repository take up 56 GB in safetensors format. License — apache-2.0. It runs via transformers. The site hosts 8 quantized builds: EXL, GGUF, MLX, NVFP4 and FP8. The most compact working build is Q4_0 at 15.9 GB. Running it locally requires a GPU with enough memory for the chosen build. Q4_0 quantization lowers the requirements compared to the full model, but the developer does not specify the minimum VRAM for specific formats. You can follow new models and builds in the CPU3D news feed.
Hy-Memory-30B-A3B: external memory for AI agents Hy-Memory-30B-A3B 30.1B parameters 56 GB safetensors Apache-2.0 full model 3B active parameters text-generation model's job 8 quantized builds EXL · GGUF · MLX NVFP4 · FP8 Q4_0 — 15.9 GB most compact External memory Mem0 MemOS Hy-Memory memory bank What the model does decides what to store how to represent data how to update memory writes to external systems Local launch GPU transformers GPU enough VRAM Track new models and builds — CPU3D news feed
How the method works. The diagram was drawn for this news note.

See also