IAAR-Shanghai

/

MemReranker-4B

like 3

Follow

Memtensor Research Group 50

Hi team — congratulations on MemReranker, the benchmark results
are impressive.

I'm building a local-first personal AI on Apple Silicon and the
0.6B variant would be ideal for my memory budget (Qwen3-14B
already runs locally). I noticed:

• MemReranker-4B has downloadable weights here, license
Apache 2.0 — clear and helpful
• MemReranker-0.6B is referenced in the paper and on the
Memos API page, but the IAAR-Shanghai/MemReranker-0.6B
HuggingFace page returns "model not found" or auth-gates
when I try to access it
• The Memos hosted API isn't an option for me (local-first
privacy constraint)

Two questions:

1. Are the 0.6B weights planned for public release on
   HuggingFace under the same Apache 2.0 license as 4B?
2. If not — would you consider it? Many local-AI builders
   would benefit from a sub-1GB agent-memory reranker, and
   it would significantly expand MemReranker adoption beyond
   teams that can spare ~10GB for the 4B variant.

Either way, thank you for the work and for open-sourcing 4B.

— Chirag M