Hy-Memory-30B-A3B has public Apache-2.0 weights, framework code and evaluation utilities.
Tencent releases a specialist writer model for long-term agent memory
2026-10-15Frontier watchTRIAL
Tencent released Apache-2.0 weights, framework code and evaluation utilities for Hy-Memory-30B-A3B. The 30B-total, roughly 3B-active model specializes in the write path for long-term memory and can replace general models inside Mem0, MemOS and Hy-Memory. On a fixed 20% sample, the authors report a small average lead over Claude Opus 5 across four memory harnesses.
Official · ReferenceHy-Memory-30B-A3B model card and weightsOfficial · ReferenceHy-Memory framework and evaluation utilities
Availability: Apache-2.0 weights, framework code and evaluation utilities
What this means for you
For long-running agent teams, assigning “what to remember” to a smaller specialist could be cheaper and more stable than calling a frontier general model. The evidence is author-run on a subsample with equal-weight aggregation, so memory growth, contradictions and retrieval outcomes still need local testing.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Model | 30B total, about 3B active; dedicated memory writer |
|---|---|
| Availability | Apache-2.0 weights, framework code and evaluation utilities |
| Evaluation | fixed 20% benchmark subsample; six equally weighted metrics across four harnesses; EverMemOS held out |
| Vendor Result | 55.49 average vs Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64 |
| Time Basis | HF repository created 2026-10-15T09:52:00Z; found in the scheduled gap review |
Technical impact 4
Engineering value 5
Adoption signal 2
Maturity 2
Verification cost 3
Risk 2