Hy-Memory-30B-A3B Text Generation Model: 12GB+ VRAM, GGUF Builds

Hy-Memory-30B-A3B Text Generation Model: 12GB+ VRAM, GGUF Builds

At a Glance

Item Value
Repository tencent/Hy-Memory-30B-A3B
Family guide Hy-Memory guide (1 articles)
Publisher guide Tencent: models and licenses
Published 2026-10-15
License apache-2.0
Formats safetensors
Paper arXiv:2610.22196
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Tencent Hunyuan has released “Hy-Memory-30B-A3B", a model specialized in writing processes for external memory frameworks. This model acts as a replacement for general-purpose LLMs in long-term external memory systems such as Mem0, MemOS, and Hy-Memory, determining what information to retain, how to represent it, and how to update the persistent memory bank in conversation, work, and coding scenarios.

According to the paper “Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing" (arXiv:2610.22196), it is designed to eliminate the increased operational costs and poor cross-protocol versatility caused by conventional frameworks calling general-purpose LLMs via prompts. It is trained in two stages: supervised distillation via Memory On-Policy Distillation (OPD), and Attributed Memory reinforcement learning (RL) with chunk-level credit assignment. It is reported that the model can transfer directly to “EverMemOS", an external framework unseen during training.

Specifications

  • Architecture: HYV3ForCausalLM (hy_v3)
  • Total Parameters: 30.1B
  • Active Parameters: Approx. 3B

Performance

In benchmark evaluations by the publisher, comparisons are conducted by replacing only the memory writing model (Memory writer) being evaluated while completely fixing the surrounding environment, such as the harness version, prompts, retrieval settings, answer generation model, and judge model. Note that “DeepSeek V4.1 Flash" is used for the comparison target DeepSeek Flash endpoint.

The macro-average scores across six metrics (LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS, and Code Memory Bench) targeting a fixed 20% subsample across the four main harnesses (Mem0, Hy-Memory, MemOS, and EverMemOS, which was excluded from the training set) are as follows:

→ Scroll horizontally to see all columns

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
Hy3 45.99 58.06 50.19 52.30 51.63
Qwen3.6-35B-A3B 47.51 55.87 47.36 50.75 50.37
Hy3-A3B Base 40.20 50.39 49.05 50.98 47.66
Qwen3-30B-A3B 39.81 45.67 45.44 46.57 44.37

As can be seen from this table, Hy-Memory-30B-A3B achieved the top average score in all four evaluated harnesses, recording an overall average (Avg.) of 55.49, surpassing Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64. In particular, it achieved 56.89 in the unseen EverMemOS harness not used during training, demonstrating high transfer performance across frameworks. It also shows a significant score gap compared to the base model Hy3-A3B Base (overall 47.66) and the similar-scale Qwen3.6-35B-A3B (overall 50.37).

Looking at the detailed breakdown data (Tables 2 to 5 in the model card), strengths and weaknesses vary depending on the domain. For example, in the Mem0 harness, it recorded the top figures among the comparison targets in LME (73.00) and Travel (54.51), while in LoCoMo (59.54) and CMB (67.13) it lagged behind Claude Opus 5 (LoCoMo 69.47, CMB 70.74). Furthermore, regarding Shopping PS (Shop.), it scored 18.33 in Mem0 (Qwen3.6-35B-A3B scored 20.00) and 21.11 in MemOS (DeepSeek V4.1 Flash scored 21.67), showing results equivalent to or slightly falling behind other models in specific tasks.

Strengths and Use Cases

Hy-Memory-30B-A3B is a model specialized in handling the “write path" to persistent memory in various external memory frameworks. It functions as a replacement for conventional general-purpose large language models in systems like Mem0, MemOS, and Hy-Memory. English and Chinese tags are assigned as supported languages.

The specific roles and areas of expertise of this model are as follows:

  • Memory Selection and Updating: Determines which information from conversation and work history should be retained as long-term memory, structures it into concise and self-contained statements, and appropriately updates the existing memory bank.
  • Adaptation to Diverse Scenarios: Supports not only regular dialogue (Chat) but also working environments handling task execution logs (Working) and coding scenarios handling repository information (Coding).
  • Cross-Framework Versatility: Possesses the flexibility to operate by transferring directly to unlearned frameworks (such as EverMemOS) without depending on prompts or rules fixed for a specific framework.
  • Reduction of Operational Costs: Compared to configurations that constantly call general-purpose frontier models, the compact design with approximately 3B active parameters significantly reduces computational resources and operational costs for the memory writing pipeline.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 30.1B parameters

Your VRAM Quantization File size Est. memory needed
12GB (RTX 4070 / 3060 12GB, etc.) IQ2_XXS 8.5GB 10.2GB
24GB (RTX 4090 / 3090, etc.) Q4_K_M 17.0GB 20.4GB
32GB (RTX 5090, etc.) Q6_K 23.0GB 27.6GB
48GB (RTX 6000 Ada / A6000, etc.) Q8_0 29.8GB 35.7GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build tencent/Hy-Memory-30B-A3B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but tencent/Hy-Memory-30B-A3B-GGUF provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compression: the IQ2_XXS build measures 2.43 bits per weight — about 15% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
LiquidAI/LFM2.5-24B-A2B-Base 24.0B 80GB lfm1.0 LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-FP8 24.0B 32GB lfm1.0 LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-MLX-bf16 24.0B 80GB lfm1.0 LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-MLX-8bit 24.0B 32GB lfm1.0 LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B 24.0B 12GB lfm1.0 LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17)

How to Get It

The weights of Hy-Memory-30B-A3B are released in safetensors format under the Apache-2.0 license. Hugging Face terms of service agreement (gated feature) is not set, allowing anyone to download and use it directly.

vLLM (confirmed working in v0.31.0) is suggested as the inference engine. You can use the model via an OpenAI-compatible API endpoint by starting the server with --served-model-name hy-memory as follows:

vllm serve tencent/Hy-Memory-30B-A3B --served-model-name hy-memory

The model card introduces an example Python script that extracts user facts and preferences using an OpenAI-compatible client with the thinking mode disabled (no-think mode).

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="hy-memory",
    messages=[
        {
            "role": "system",
            "content": (
                "Extract explicit user facts and preferences from the conversation. "
                "Write each memory as a concise, self-contained statement. "
                "Do not infer or invent information. "
                'Return only a JSON object with a "memories" list of strings.'
            ),
        },
        {
            "role": "user",
            "content": (
                "My name is Alex. I live in Seattle and work as a software engineer. "
                "I mainly use Python and prefer concise answers with code examples. "
                "I'm vegetarian, so please keep that in mind when suggesting meals."
            ),
        },
    ],
    temperature=0.0,
    max_tokens=512,
    extra_body={
        "chat_template_kwargs": {
            "reasoning_effort": "no_think",
        },
    },
)

Additionally, the framework integration code and a complete set of evaluation utilities are published in the GitHub repository Tencent-Hunyuan/Hy-Memory.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-18 tencent FP8 tencent/Hy-Memory-30B-A3B-FP8 FP8 34.2GB (fits in 48GB VRAM)
2026-10-19 tencent GGUF (imatrix) tencent/Hy-Memory-30B-A3B-GGUF IQ2_XXS 10.2GB (fits in 12GB VRAM)
2026-11-17 mlx-community MLX mlx-community/Hy-Memory-30B-A3B-4bit MLX 4bit 18.9GB (fits in 24GB VRAM)
2026-11-17 mlx-community MLX mlx-community/Hy-Memory-30B-A3B-8bit MLX 8bit 35.7GB (fits in 48GB VRAM)

File sizes of each build:

  • Available builds in tencent/Hy-Memory-30B-A3B-FP8: FP8 28.5GB
  • Available builds in tencent/Hy-Memory-30B-A3B-GGUF: IQ1_M 7.5GB / IQ2_XXS 8.5GB / Q3_K_M 13.5GB / Q4_K_M 17.0GB / Q6_K 23.0GB / Q8_0 29.8GB
  • Available builds in mlx-community/Hy-Memory-30B-A3B-4bit: MLX 4bit 15.8GB
  • Available builds in mlx-community/Hy-Memory-30B-A3B-8bit: MLX 8bit 29.8GB

In addition, 20 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-17: Added converted builds to “Quantized and Converted Variants”: mlx-community/Hy-Memory-30B-A3B-4bit, mlx-community/Hy-Memory-30B-A3B-8bit
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of tencent/Hy-Memory-30B-A3B-GGUF.
  • 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).