Apache-2.0llmRuns locally30B
Hy-Memory: Specs, License & How to Run
Tencent Hunyuan memory-writer MoE (30B-A3B) plus Hy-Memory SDK and MemHub for agent memory.
- License
- Apache-2.0
- Parameters
- 30B (active 3B)
- HF likes
- 188 (2026-10-19)
- First seen
- 2026-10-15
TL;DR. Hy-Memory is a llm model from Tencent Hunyuan. Parameters: 30B (active 3B).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs, Cursor, Mem0.
| Developer | Tencent Hunyuan |
|---|---|
| Type | llm |
| Architecture | HYV3 MoE (config num_experts 128, num_experts_per_tok 8); memory-writer for external memory frameworks |
| Parameters | 30B · active 3B |
| Modalities | text |
| Languages | en, zh |
| License | Apache-2.0 — Apache-2.0 on the Hugging Face card (license apache-2.0) and on the GitHub LICENSE file (“Hy-Memory is licensed under Apache-2.0”) as of 2026-10-19. GitHub’s license API field is the opaque label “other”; we copy the LICENSE file. Not legal advice. |
| Hardware | Not yet verified — Card quickstart deploys with vLLM 0.31.0 as `--served-model-name hy-memory`. HF API safetensors.parameters total 30,064,725,888 (~30.06B); usedStorage 60,132,268,315 bytes (~56 GiB on disk). Peak VRAM is not independently measured. (not independently tested) |
| Hugging Face likes | 188 (as of 2026-10-19) |
| Model card | https://huggingface.co/tencent/Hy-Memory-30B-A3B |
| First seen | 2026-10-15 |
Hy-Memory (Hy Memory 2.0) is Tencent Hunyuan’s open memory stack: a 30B-A3B writer model, the hy-memory Python SDK, and MemHub. Official hosts we opened on 2026-10-19: tencent/Hy-Memory-30B-A3B, github.com/Tencent-Hunyuan/Hy-Memory (779 stars), the LICENSE file, the product page memory.hunyuan.tencent.com (dated 2026.10.16), and the technical report arXiv 2610.22196.
This page hangs off ChatGPT alternatives, closed LLM APIs, and Cursor alternatives because MemHub’s README names Cursor (and other coding agents) as apps it can import history from. Those mappings are editorial. This is not a ChatGPT or Cursor quality ranking.
What we verified (2026-10-19)
| Field | Value | Source |
|---|---|---|
| HF likes / downloads | 188 / 1,187 | HF API tencent/Hy-Memory-30B-A3B |
| Card created / last modified | 2026-10-15T09:52:00Z / 2026-10-16T08:12:51Z | same |
| Gated | false | HF API |
| License (card) | apache-2.0 | HF cardData |
| LICENSE file | Apache-2.0 (Tencent preamble + Apache 2.0 terms) | GitHub LICENSE |
| GitHub license API | other | GitHub API (we still copy the file) |
| Pipeline tag | text-generation | HF API |
| Languages | en, zh | HF cardData |
| safetensors total | 30,064,725,888 (~30.06B) | HF API |
| usedStorage | 60,132,268,315 bytes (~56 GiB) | HF API |
| Config experts | 128 experts, 8 per token | HF config |
| GitHub stars / forks | 779 / 66 | GitHub API |
| GitHub created / last push | 2026-10-15T13:06:56Z / 2026-10-18T14:30:24Z | same |
| MemHub standalone repo | 404 | GitHub API Tencent-Hunyuan/MemHub |
The card describes a compact long-term-memory writer (~3B active parameters) trained in two stages: Memory On-Policy Distillation, then Attributed Memory Reinforcement Learning. It is meant to replace a general-purpose LLM on the write path inside Mem0, MemOS, and Hy-Memory (Chat / Working / Coding scenes) and is claimed to transfer to EverMemOS (held out from training). Those sentences are upstream.
The product site’s three layers, copied as labels only: Hy-Memory-30B-A3B (model), Hy-Memory SDK (engine), MemHub (app). GitHub README: MemHub is a local-first desktop and CLI that unifies conversation history and coding traces from user-authorized apps (examples named: Claude Code, Cursor, Codex, OpenCode, Qwen Code).
How to run (from official README, not yet tested)
Weights: serve tencent/Hy-Memory-30B-A3B with vLLM 0.31.0 and --served-model-name hy-memory, then call the OpenAI-compatible API. The card’s example uses reasoning_effort: "no_think" via chat_template_kwargs.
SDK: GitHub news names Hy Memory 2.0 (hy-memory==2.0.4) on PyPI. Install snippets on the repo still also show pip install hy-memory==2.0.3. Point MEMORY_LLM_BASE_URL at the writer endpoint; pro mode uses the LLM, lite does not. forkalt has not installed the package.
License
Apache-2.0 on the HF card and the GitHub LICENSE file (“Hy-Memory is licensed under Apache-2.0”). GitHub’s license API key is other because of the Tencent preamble; we still record SPDX Apache-2.0 from the file. Not legal advice.
What we did not verify
Table 1 on the card (Hy-Memory-30B-A3B overall 55.49 vs Claude Opus 5 54.62 vs DeepSeek V4.1 Flash 53.64 on a fixed 20% subsample across Mem0 / Hy-Mem. / MemOS / EverMemOS) stays self-reported. Evaluation notes say every “DeepSeek Flash” endpoint in that report denotes DeepSeek V4.1 Flash. We have not reproduced the matrix. Peak VRAM, MemHub macOS/Linux runtime artifacts, and the 404 standalone MemHub repo are not filled in with guesses.
Related: Hy-Nomas (same Hunyuan org; interactive video world model, not a memory writer), Hy-SimulMT (same Hunyuan 30B-A3B line; simultaneous speech MT, not a memory writer), Hindsight on Mem0 alternatives (MIT memory service, not a writer model), DeepSeek-V4.1-Flash (the writer-baseline named on the Hy-Memory card), vLLM (the card names vLLM 0.31.0), DeepInfer, DeepInfer vs vLLM, Reflection Beam (coding/agent weights, different job), ChatGPT alternatives.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Cross-harness macro average (20% subsample, vendor) | 55.49 (self-reported) | https://huggingface.co/tencent/Hy-Memory-30B-A3B |
Deploy options
FAQ
Is Hy-Memory a chatbot with built-in memory?
No. The 30B-A3B checkpoint is a memory writer: it decides what to store and how to update an external memory bank (Mem0, MemOS, Hy-Memory, or a held-out harness such as EverMemOS). Persistent memory still lives in the framework / vector store you run. MemHub is the desktop/CLI product that uses the SDK across authorized AI tools.
Are the weights and SDK open?
Yes. huggingface.co/tencent/Hy-Memory-30B-A3B is ungated (created 2026-10-15T09:52:00Z; 188 likes / 1,187 downloads in the 2026-10-19 API check; card license apache-2.0). github.com/Tencent-Hunyuan/Hy-Memory has 779 stars / 66 forks (created 2026-10-15; last push 2026-10-18). Official site dated 2026.10.16 describes Hy-Memory 2.0 as model + engine + MemHub. GitHub news line names `hy-memory==2.0.4` on PyPI.
Where is MemHub’s GitHub repo?
The product site links github.com/Tencent-Hunyuan/MemHub. GitHub API returned 404 for that path on 2026-10-19. CONTRIBUTING.md in Tencent-Hunyuan/Hy-Memory says the repository also contains the MemHub application under memhub/. We do not invent a second repo that 404s.
Does it replace ChatGPT memory?
We do not claim that. This page identifies a memory-specialized writer MoE and an open SDK. Vendor tables versus Claude Opus 5 and DeepSeek V4.1 Flash stay self-reported. There is no sourced bake-off against ChatGPT memory.
Sources & verification
- https://huggingface.co/tencent/Hy-Memory-30B-A3B
- https://github.com/Tencent-Hunyuan/Hy-Memory
- https://github.com/Tencent-Hunyuan/Hy-Memory/blob/main/LICENSE
- https://memory.hunyuan.tencent.com/
- https://arxiv.org/pdf/2610.22196
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.