Hy-Memory-30B-A3B: novidades, hardware necessário, como rodar grátis

Tencent’s Hy-Memory-30B-A3B packs a massive 256 K token context into a compact writer model. It runs with only about 3 B active parameters per token, making long-term memory feasible at home.

Hy-Memory-30B-A3B is a writer model created by Tencent Hunyuan. The full model contains 30.1 billion parameters, but only roughly three billion are active for each token thanks to a mixture-of-experts design. It can attend to a context of 256 K tokens, far larger than typical LLMs. The weights are released under an Apache 2.0 license, so anyone can download them from Hugging Face. Because of its expert routing, the model can fit on a single high-end GPU when quantized to 4-bit, using about 21 GB of memory for an 8 K context. This makes it one of the most capable open-source options for long-term memory tasks.

According to the model card, the model is trained in two distinct stages. First, Memory On-Policy Distillation provides dense teacher supervision on state chains that the model writes. Then, an Attributed Memory Reinforcement Learning phase evaluates the resulting memory banks, assigns credit at the chunk level, and carries the best bank forward across windows. The same harness configurations are used in both stages, ensuring the writer learns directly in the environments where it will be deployed. This training scheme lets Hy-Memory-30B-A3B generalise to memory frameworks it has never seen, such as EverMemOS, without additional fine-tuning.

The model card lists a cross-harness macro-average of 55.49 on a fixed 20 % benchmark subsample. Across the four harnesses, Mem0, Hy-Memory, MemOS, and EverMemOS, it achieves the highest overall average among the compared writers. Its strongest single harness is Hy-Memory, where it scores 60.55, slightly ahead of Claude Opus 5’s 60.53. On the held-out EverMemOS harness, Hy-Memory-30B-A3B records 56.89, again beating the next best model. These numbers come from six equally weighted metrics per harness, covering judgment, long-term memory evaluation, persona memory, travel, shopping, and code memory benchmarks.

The results show the model’s greatest strength is its ability to maintain and update a persistent memory bank, especially in the Hy-Memory harness where it leads all competitors. It also performs well on long-term memory and code memory benchmarks. However, its shopping scenario score is comparatively lower, indicating a weaker grasp of that domain. To run the full 256 K context the model must be quantized to 4-bit, which reduces memory usage to about 21 GB for an 8 K window. This quantization is a requirement for most consumer-grade GPUs, and it may affect precision on some tasks.

If you have a GPU like an RTX 3090, the model fits entirely in GPU memory when run at 4-bit precision, making it practical for personal machines. Developers building applications that require very long context, such as multi-turn chat, document analysis, or codebases, will find the 256 K token window valuable. Researchers exploring memory-augmented language models can use the two-stage training pipeline as a reference point. Finally, anyone who prefers fully open-source models under a permissive Apache 2.0 license can download the weights from Hugging Face and experiment without licensing hurdles.

O que você precisa para rodá-lo em casa? Com 4 bits, ele ocupa cerca de 21 gigabytes de memória. A menor máquina doméstica que o comporta por inteiro é RTX 3090. Em RTX 3060 12 GB com 32 gigabytes de RAM, ele escreve cerca de 52 tokens por segundo. Em RTX 4090 com 64 gigabytes de RAM, ele escreve cerca de 163 tokens por segundo.

Ainda não há API grátis para ele. Os pesos são abertos: você pode baixá‑los do Hugging Face sob a licença Apache 2.0. A Arena ainda não o classificou: adicionaremos a pontuação assim que os votos chegarem.

Compare Hy Memory 30B A3B with every free model, and check what your own GPU can run, on FreeLLMArena.