Tencent Releases Hy-Memory-30B-A3B: Open-Weight Model Claims Average Benchmark Lead Over Claude Opus 5

Tencent's Hunyuan team released the open-weight model Hy-Memory-30B-A3B on Hugging Face on October 16. Designed specifically for AI agent long-term memory management, the model determines which information to retain from conversations and how to update existing memories.

According to PC Watch, the model utilizes a Mixture-of-Experts (MoE) architecture with 30 billion total parameters but only activates approximately 3 billion during inference, aiming to reduce deployment costs. Tencent's internal benchmarks indicate that across four major memory frameworks including Mem0 and MemOS, Hy-Memory-30B-A3B achieved an average score of 55.49, slightly surpassing Anthropic's Claude Opus 5 at 54.62.

However, this advantage is not uniform. In the LoCoMo conversational memory test within the Mem0 framework, Hy-Memory-30B-A3B scored 59.54, significantly trailing Claude Opus 5's 69.47. On the Hy-Memory framework itself, the scores were nearly identical (60.55 vs. 60.53).

The model is released under the Apache 2.0 license, permitting free commercial use, modification, and redistribution. Its training involved two stages: first distilling knowledge from a larger teacher model via 'Memory On-Policy Distillation,' followed by reinforcement learning ('Attributed Memory RL') to optimize memory write accuracy. These performance figures are self-reported by Tencent and have not yet been independently reproduced.