What is Hy-Memory-30B-A3B?
Hy-Memory-30B-A3B is a mixture-of-experts model from Tencent Hy (formerly Hunyuan), released on Hugging Face on 15 October 2026 with a technical report on arXiv. It has about 30 billion parameters, 3 billion active per token, and is built on Tencent's Hy3-A3B base.
It is not a chat model. It replaces the general-purpose LLM that memory frameworks such as Mem0, MemOS and Tencent's own Hy-Memory call to extract facts from conversations, decide what to keep and update the stored memory bank. It was trained in two stages, on-policy distillation from Tencent's Hy3 flagship and then reinforcement learning that scores each chunk of memory it writes.
The weights are under Apache 2.0, as a single BF16 checkpoint of about 60GB. Tencent published no launch post and there is no hosted API.
How much does Hy-Memory-30B-A3B cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0, BF16 only |
| Hosted API | Not offered |
Open weights under Apache 2.0; no hosted API.
Is Hy-Memory-30B-A3B worth it?
Hy-Memory-30B-A3B targets a cost most agent builders notice late: every conversation turn sent to Mem0 or a similar framework triggers another call to an expensive model just to decide what to remember. Tencent's answer is a 3B-active specialist, and on its own evaluation it matches the frontier. Across four memory frameworks it averages 55.49, against 54.62 for Claude Opus 5, 53.64 for DeepSeek V4.1 Flash and 50.37 for Qwen3.6-35B-A3B.
The lead is thin and the method is limited. On the Hy-Memory and MemOS harnesses the gap to Opus is under a tenth of a point; the separation comes from Mem0 (0.6 points) and the held-out EverMemOS harness (2.8 points). Every benchmark ran on a fixed 20% subsample, and Opus still wins on several individual metrics, including LoCoMo inside Mem0 (69.47 to 59.54). All writers score between about 14 and 24 on the shopping-preference test, so preference memory remains weak whichever model writes it.
The strongest evidence is transfer. EverMemOS was held out of training, and the model's top score there (56.89) suggests it learned the job rather than one framework's prompts. That matters because memory-writing prompts differ a great deal between frameworks.
Deployment is less polished than the research. There is one BF16 checkpoint of about 60GB in 51 shards, with no FP8, GGUF or quantised builds, so it needs an 80GB-class GPU or two smaller ones, far more than its 3B active parameters suggest. For a team already paying a frontier API for memory writes at scale, that hardware is the price of the swap.
Where it earns its place
A backend engineer at a companion-app startup with 500,000 daily users
Cut the API bill for Mem0 memory extraction
Point Mem0's LLM at a vLLM server running hy-memory in no-think mode. Every user turn that would have gone to a frontier model now runs on a self-hosted 3B-active model; the answering model stays unchanged.
An ML engineer building a coding assistant with persistent project memory
Record repository facts across sessions
Tencent's Code Memory Bench scores inside Mem0 were 67.13 against 70.74 for Opus, so it trails slightly here; test on your own repositories before switching.
A researcher comparing agent memory frameworks
Hold the writer constant across Mem0, MemOS and EverMemOS
The released evaluation utilities on GitHub fix retrieval, prompts and judge so only the writer changes, which gives a reproducible baseline for new memory designs.
Before you commit
- Setup
- Serve with vLLM (tested with v0.31.0) as hy-memory and call the OpenAI-compatible API with reasoning_effort set to no_think, as in Tencent's quickstart; then set it as the LLM in your memory framework.
- Integrations
- Mem0, MemOS, Tencent's Hy-Memory framework (chat, working and coding scenarios); shown to transfer to EverMemOS; framework code and evaluation tools on GitHub.
- Data and privacy
- Apache 2.0 with no use restrictions. Self-hosted, so user memories never leave your infrastructure, which is the main privacy gain over sending every turn to a hosted model.
- Cost at scale
- About 60GB of BF16 weights, so one 80GB H100 or two 48GB GPUs; there are no quantised builds yet to fit smaller cards.
- Limits
- Writes memory only; English and Chinese; context length not stated on the card; benchmark results not independently reproduced.
What’s good
- Matches Claude Opus 5 as a memory writer on Tencent's four-framework average (55.49 to 54.62)
- Best score on the held-out EverMemOS harness, so it transfers between frameworks
- Drop-in for the LLM inside Mem0 and MemOS
- Apache 2.0
What’s not
- All results from Tencent, on 20% benchmark subsamples
- Lead over Opus under 0.1 points on two of four harnesses
- Single 60GB BF16 checkpoint; no quantised builds
- Does one job only; you still need a separate answering model
Best for
Teams running Mem0, MemOS or similar memory layers at volume who want to replace a paid frontier model on the write path with self-hosted weights.
Hy-Memory-30B-A3B alternatives
Hy-Memory-30B-A3B: common questions
- Is Hy-Memory-30B-A3B worth it?
- A narrow model with a sound idea, writing agent memory as well as Claude Opus 5 on Tencent's tests at a fraction of the cost, though only Tencent has measured it. We rate it 3.7 out of 5.
- How much does Hy-Memory-30B-A3B cost?
- Open weights under Apache 2.0; no hosted API.
- What are the downsides of Hy-Memory-30B-A3B?
- All results from Tencent, on 20% benchmark subsamples; lead over Opus under 0.1 points on two of four harnesses; single 60GB BF16 checkpoint; no quantised builds; does one job only; you still need a separate answering model.
- Who is Hy-Memory-30B-A3B best for?
- Teams running Mem0, MemOS or similar memory layers at volume who want to replace a paid frontier model on the write path with self-hosted weights.
- How much does Hy-Memory-30B-A3B cost for a team or in production?
- About 60GB of BF16 weights, so one 80GB H100 or two 48GB GPUs; there are no quantised builds yet to fit smaller cards.
- Is Hy-Memory-30B-A3B safe for company data?
- Apache 2.0 with no use restrictions. Self-hosted, so user memories never leave your infrastructure, which is the main privacy gain over sending every turn to a hosted model.
- What are the best Hy-Memory-30B-A3B alternatives?
- The closest alternatives we have reviewed are Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash.
Answers checked Oct 31, 2026.