Description
Hy-Memory-30B-A3B is a sophisticated long-term-memory writer model created by Tencent Hunyuan. Designed to operate within external memory frameworks, it replaces general-purpose LLMs to enhance the efficiency of information retention and representation. The model is particularly effective in Chat, Working, and Coding scenarios, where it decides what information to retain and how to represent it accurately. It is trained in two stages: Memory On-Policy Distillation (OPD) and Attributed Memory Reinforcement Learning. This training allows it to transfer effectively to unseen memory frameworks such as EverMemOS.
Read More
The model's performance is evaluated using a fixed 20% benchmark matrix, where it demonstrates superior capabilities across various metrics. It achieves high scores in harnesses like Mem0, Hy-Memory, MemOS, and EverMemOS, indicating its robustness and adaptability. The evaluation process ensures that the memory writer's efficiency is maximized by fixing the harness version, prompts, and retrieval configuration while keeping other framework-side components constant.
Hy-Memory-30B-A3B's training involves dense teacher supervision on student-written state chains during the OPD stage, followed by reinforcement learning that assigns chunk-level credit and carries selected banks across windows. This two-stage training process ensures that the model can adapt to different scenarios and frameworks effectively.
The model is available for deployment with vLLM and can be accessed via an OpenAI-compatible API. This makes it accessible for developers looking to integrate advanced memory writing capabilities into their applications. The model's weights, configuration, and technical report are available on Hugging Face, providing comprehensive resources for further exploration and utilization.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentvLLM
- PlatformsWeb
Hy-Memory-30B-A3B Highlights
Compact long-term-memory writer model
Developed by Tencent Hunyuan
Replaces general-purpose LLMs
Operates in Chat, Working, and Coding scenarios
Two-stage training: OPD and reinforcement learning
Transfers to unseen memory frameworks
Evaluated on fixed 20% benchmark matrix
Available via OpenAI-compatible API
Getting Started with Hy-Memory-30B-A3B
Access model page: Visit the Hugging Face model page
Authenticate: Ensure you have the necessary credentials
Configure: Set up the model parameters as needed
Send prompts: Use the API to send prompts to the model
Fine-tune: Adjust the model for specific use cases if required
Hy-Memory-30B-A3B's Use Cases
- Chat Optimization
- Coding Assistance
- Information Retention
- Framework Adaptation
- Benchmark Evaluation
