Tencent releases Hy-Memory 30B-A3B under Apache 2.0

Tencent has published Hy-Memory 30B-A3B, a text-generation language model built around a mixture-of-experts (MoE) design. According to the model's Hugging Face repository, it carries 30 billion total parameters but activates only about 3 billion per token — the "A3B" in its name — a configuration meant to deliver the capacity of a larger model at a fraction of the inference cost.

The standout framing here is memory. As the name suggests, the model is positioned around memory-centric capabilities, a growing area of interest as developers look to build assistants that retain context across longer interactions rather than treating every exchange in isolation.

Why it matters

Sparse MoE architectures have become the default strategy for teams trying to balance quality and serving cost, and a 30B/3B split puts Hy-Memory in a practical range for self-hosting.

• Open license: Apache 2.0, which permits commercial use and modification.
• Efficient inference: roughly 3B active parameters per token keeps compute requirements modest.
• Focused design: the model emphasizes memory rather than general-purpose breadth.

Tencent has steadily expanded its open-weights footprint, and an Apache-licensed release lowers the barrier for researchers and companies that want to experiment without restrictive terms. Full details on context length, benchmarks, and recommended usage are available on the official repository.