Tencent releases Hy-Memory-30B-A3B for external memory frameworks
Tencent Hunyuan has released Hy-Memory-30B-A3B, a lightweight model (~3B active parameters) designed to manage long-term memory in external frameworks like Mem0, MemOS, and Hy-Memory. It replaces general-purpose LLMs in these systems, deciding what data to store, how to represent it, and how to update persistent memory banks. The model was trained in two stages—Memory On-Policy Distillation…
Key points
- Hy-Memory-30B-A3B (~3B params) replaces general-purpose LLMs in memory frameworks like Mem0 and MemOS
- Trained via OPD and reinforcement learning, tested on EverMemOS (held-out benchmark)
- Deployable via vLLM (v0.31.0) with OpenAI-compatible API; weights on Hugging Face
Tencent provides deployment instructions via vLLM (v0.31.0) with an OpenAI-compatible API. The model is available on Hugging Face with weights, configuration, and a model card. Benchmarking details, including six metrics (LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS, and Code Memory Bench), are documented in an accompanying arXiv report. The framework code and utilities are open-source on GitHub.
Model page: DeepSeek-V4.1-Flash →
tencent/Hy-Memory-30B-A3B · Hugging Face
huggingface.co · 16 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Nvidia pledges $250M in compute for UK AI agent safety research · 9 src
- Linus Torvalds posts first AI-assisted Linux kernel patch with 23x benchmark gain · 2 src
- PhD students using AI in dissertations less likely to enter academia · 1 src
- FPT expands Nemotron-Personas-Vietnam dataset to 1M personas with NVIDIA · 1 src
- Llama.cpp merges AVX2 tiled accumulator reduction update · 1 src
Comments
via GitHub Discussions