Hy-Memory-30B-A3B
Tencent Hunyuan (115 likes on HF; repo live 2026-10-15; tech report arXiv:2610.22196; README captured in full): a compact long-term-memory writer — ~3B active parameters (30B-A3B MoE) that executes the write path inside external memory frameworks. It directly replaces the general-purpose LLM inside Mem0, MemOS, and Hy-Memory, deciding what to retain, how precisely to represent it, and how new evidence updates the persistent bank. Not a chatbot. Two-stage training: Memory On-Policy Distillation (dense teacher supervision on student-written state chains) then Attributed Memory RL (examines resulting banks, chunk-level credit, carries the selected bank across windows) — and it transfers zero-shot to unseen frameworks (EverMemOS held out from training).
The benchmark table (fixed 20% subsample; harness + prompts + retrieval + answer model + judge all held fixed; only the writer swaps):
| Memory writer | Mem0 | Hy-Mem. | MemOS | EverMem. | Avg. |
|---|---|---|---|---|---|
| Hy-Memory-30B-A3B | 54.02 | 60.55 | 50.51 | 56.89 | 55.49 |
| Claude Opus 5 | 53.43 | 60.53 | 50.50 | 54.04 | 54.62 |
| DeepSeek V4.1 Flash | 49.73 | 60.20 | 49.17 | 55.47 | 53.64 |
| Qwen3.6-35B-A3B | 47.51 | 55.87 | 47.36 | 50.75 | 50.37 |
A 3B-active specialist ≥ Opus 5 at the remember/forget/update judgment — and EverMemOS (the unseen harness) still ranks the Tweet’s “EverMemOS #1 at 56.89” as Hy-Memory-the-writer inside EverMemOS, i.e. the specialist wins even in a framework it wasn’t trained on.
Serving: vLLM + SGLang, OpenAI-compatible API. Community quants already out: Q4_K_M ~17 GB, Q5_K_M ~20 GB, Q6_K ~23 GB, EXL3 4.08bpw ~15 GB — TeksEdge’s framing (259❤/11.5k views; tweet captured): “roughly the same memory class as a Qwen3.8-27B Q4, except Qwen is a dense generalist and this activates ~3B/token trained specifically for remembering” — his pick to pair with Qwen3.8-27B as the memory companion.
The framework side lives at Hy-Memory (OSS) (316★): full memory framework (MemHub local-first desktop/CLI, graph memory with projection/rerank, coding/working scenes, OpenClaw + agent plugins under plugins/).
Why it matters
The specialist-beats-frontier pattern (Small language models, Decision models) arriving at the memory lane — the exact slot this vault’s Agent knowledge systems note tracks: most memory systems burn a general-purpose LLM on write-path judgment; Tencent trained a sparse MoE for that job only. Ops-relevant: 15–17 GB quant slots beside a Qwen3.8-27B on one 3090-class card — the homelab memory-writer lane is now a model-swap, not a service install.