← All models

Hy-Memory-30B-A3B

Published Publisher Tencent

Overview

Hy-Memory-30B-A3B is Tencent Hunyuan’s specialized long-term memory writer for external memory systems. It is a 30-billion-parameter mixture-of-experts model with approximately 3 billion active parameters per token. It is designed to decide what information to retain, how to represent it, and how new evidence should update a persistent memory bank. (huggingface.co)

Capabilities and uses

  • Can replace the general-purpose model used for memory writing in Mem0, MemOS and Hy-Memory, including chat, work and coding scenarios.
  • Was trained using Memory On-Policy Distillation followed by attributed reinforcement learning. Tencent also evaluated it on EverMemOS, a memory framework held out from training.
  • Tencent reports a 55.49% overall average across four memory frameworks and six equally weighted metrics, compared with 54.62% for Claude Opus 5 as a memory writer under the same evaluation. The main matrix used fixed 20% benchmark subsamples; results should be interpreted within that setup. (huggingface.co)

Availability and limitations

Tencent publishes model weights and a model card under the Apache 2.0 license. The model is specialized for the memory-writing path, not presented as a general-purpose replacement for an assistant or a complete memory system. Benchmark results do not guarantee performance on other workloads; persistent memories should be reviewed and handled with appropriate privacy controls. (huggingface.co)

Release date

The associated research paper was submitted on October 15, 2026. This confirms the paper’s public submission date, not necessarily the exact date on which the model weights first became available. (arxiv.org)

Pricing

As checked on November 16, 2026, no model-specific official hosted API input or output price was verified in Tencent Cloud’s mainland CNY price documentation or international USD price documentation. This does not establish that hosted use is free. Self-hosting the published weights can still incur infrastructure costs. (cloud.tencent.com)

Official sources