Tencent Hunyuan released Hy-Memory-30B-A3B on 2026-10-15 with a 262,144-position context and 128 experts, 8 active per token. It suits teams already running Mem0, MemOS or Hy Memory that want a small local model to write memories instead of calling a frontier API.
Can you use it?
- Free plan: not published.
- Open weights: released (Apache-2.0).
Checked 2 Nov against Tencent's own pages. Compare access across frontier models
What changed
- · Release Hy-Memory-30B-A3B weights released under Apache 2.0 source
Hy-Memory-30B-A3B is Tencent Hunyuan's open-source memory-writer model: 30 billion total parameters, about 3 billion active, Apache 2.0 licensed. It decides what an agent's long-term memory keeps, and Tencent reports a 55.49 four-framework average against 54.62 for Claude Opus 5 on its own benchmark subsamples.
Provider: Tencent · Family: Hunyuan
Context window: 262,144 tokens
Input modalities: text · Output: text
About Hy-Memory-30B-A3B
Hy-Memory-30B-A3B is a memory-writer model, not a chat assistant: its job is to read an agent's conversation or coding trace and decide what long-term memory keeps, how precisely it is worded, and how new evidence updates what is already stored. Tencent's Hunyuan team put the weights on Hugging Face on 2026-10-15 under Apache 2.0, together with a technical report (arXiv 2610.22196) titled Attributed Reinforcement Learning for Long-Term Memory Writing. The model is a mixture of experts with 30 billion total parameters and roughly 3 billion active per token: 48 layers, 128 routed experts with 8 chosen per token plus 1 shared expert, a 262,144-position context limit in its config, and a 120,832-token vocabulary. It is in the same Hunyuan lineage as Hy3 and Hy4 Preview, and a Hy3-A3B Base checkpoint is the baseline it was trained from in Tencent's comparison table.
It is meant to replace the general-purpose LLM inside the write path of a memory framework. The model card names three: Mem0, MemOS and Tencent's own Hy Memory, the last covering chat, working and coding scenes. Training has two stages. Memory on-policy distillation gives dense teacher supervision on memory banks the student wrote itself; attributed memory reinforcement learning then scores the resulting banks chunk by chunk and carries the best bank forward. Tencent holds out a fourth framework, EverMemOS, from training to test whether the skill transfers.
The benchmark evidence is vendor-run and should be read that way. Tencent evaluated every writer on fixed 20% subsamples of six measures (LoCoMo judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and Code Memory Bench), averaged per framework and then across the four frameworks. On that average the model card reports 55.49 for Hy-Memory-30B-A3B against 54.62 for Claude Opus 5, 53.64 for DeepSeek V4.1 Flash, 51.63 for Hy3 and 50.37 for Qwen3.6-35B-A3B. The gaps are small. On the Hy Memory and MemOS frameworks the model leads Claude Opus 5 by 0.02 and 0.01 points, on Mem0 by 0.59, and the clearest margin, 2.85 points, is on the held-out EverMemOS framework. It does not win everywhere: on the Mem0 framework Claude Opus 5 scores 69.47 on LoCoMo against 59.54, and DeepSeek V4.1 Flash is ahead on EverMemOS LongMemEval. We found no independent reproduction of these runs.
To run it yourself, the card shows BF16 and F32 safetensors, English and Chinese as listed languages, and serving through vLLM (tested with v0.31.0) or SGLang behind an OpenAI-compatible endpoint. Tencent also published FP8 and GGUF builds. No hosted inference provider serves it, so there is no per-token price to quote; the cost is the GPU you run it on. The quickstart asks for a JSON object holding a list of memory strings, uses temperature 0, and switches off visible reasoning with a no_think setting in the chat template.
The model is one of three layers Tencent shipped under the Hy Memory name. The Python SDK, hy-memory on PyPI, provides the chat, working and coding memory pipelines, with ChromaDB as the default vector store (Qdrant and FAISS are supported) and Kuzu or Neo4j for the graph store in its ultra mode. MemHub is a local-first desktop and command-line product built on the SDK. It gathers conversation history and coding traces from AI apps the user authorises, the README naming Claude Code, Cursor, Codex, OpenCode and Qwen Code, into one local SQLite store that supported tools can then recall from. MemHub 1.0.6 (November 2026) ships a macOS arm64 app and Linux x64 and arm64 runtimes, Windows is not supported, and the README still calls the product a preview. The repository also carries plugins for Hermes Agent and OpenClaw.
Pick it if you already run Mem0, MemOS or Hy Memory, want memory writing to stay on your own hardware, and can host a 30B-parameter model. Look elsewhere if you need a general assistant, a hosted API with an SLA, or Windows support for MemHub; a stateful agent platform such as Letta or a frontier API model will serve those needs better.
Pricing
No hosted API price is published. The weights are free to download, and the cost is the GPU you run them on. Compare memory layers such as [Mem0](/hub/tools/mem0) before choosing where the writer runs.
Key Features
- Memory-writer role: Replaces the general-purpose LLM in the write path of Mem0, MemOS and Hy Memory, deciding what to keep and how to update it.
- Two-stage training: Memory on-policy distillation first, then attributed memory reinforcement learning that scores each memory chunk.
- Small active size: Sparse routing activates about one tenth of the weights per token, roughly 3B of 30B.
- 262,144-position context: The config sets max_position_embeddings to 262144, enough for long conversation or coding transcripts.
- Open quantized builds: Official FP8 and GGUF versions sit beside the BF16 weights on Hugging Face.
Pros
- Matches Claude Opus 5 on a four-framework memory average (55.49 vs 54.62) in Tencent's tests.
- Apache 2.0 weights with FP8 and GGUF builds allow fully local memory writing.
- About 3B active parameters keeps serving cost near that of a small model.
Cons
- Benchmarks are vendor-run on 20% subsamples and unreproduced.
- Only useful as a memory writer; it is not a general assistant.
- No hosted provider, so you must run a 30B-parameter model yourself.
Frequently Asked Questions
What does it cost to run Hy-Memory-30B-A3B?
No hosted API price exists: the Hugging Face card lists no inference provider, and the weights are free under Apache 2.0. The cost is your own GPU time. At BF16 the 30B weights alone come to roughly 60 GB by simple arithmetic (2 bytes per parameter), before cache, which is why Tencent also published FP8 and GGUF builds.
Is it better than a frontier model at writing agent memory?
Tencent's own tables put it marginally ahead of Claude Opus 5 on a four-framework average, and the gap on three of the four frameworks is under 0.6 points. Treat that as parity at a far smaller active size, not a win. Claude Opus 5 is clearly ahead on LoCoMo inside the Mem0 framework, and nobody outside Tencent has reproduced the runs.
Can I use Hy-Memory-30B-A3B commercially, and where do I download it?
The weights are open source under Apache 2.0, which permits commercial use with the usual notice terms. Download them from the tencent organisation on Hugging Face; FP8 and GGUF variants sit beside the BF16 original. The SDK and MemHub are also Apache 2.0.
Does MemHub send my conversations to Tencent?
The README describes MemHub as local-first: authorised conversation and coding history is merged into one local store under your home directory, and the model provider you configure does the processing. Source reading, model processing and host reading are separate permissions you grant. The README says the product captures no screen, OCR, clipboard or microphone data. Read PRIVACY.md in the repository before pointing it at a hosted model.
Who should skip Hy-Memory-30B-A3B?
Skip it if you want a chat assistant, because it is trained only to write memories and expects a system prompt asking for a JSON memory list. Teams without a GPU large enough for a 30B mixture of experts, or on Windows, will have a harder time with MemHub than with a hosted memory layer like Mem0's cloud.