Hy-Memory-30B-A3B
An open-source task-oriented MoE model from Tencent trained specifically as a Memory Provider. It has 30 billion total parameters with roughly 3 billion activated during inference. It is proficient at memory management tasks such as generating conversation summaries, context compression, and entity recognition and writing, and it produces text output rather than latent representations.
Timeline (last 90 days)
On the MLX side, upstream mlx-lm does not yet recognize hy_v3, but hermitdave's Mixed4-6 build bundles hy_v3.py and points to it via config, so it can be loaded directly with stock mlx-lm; the uploader reports decoding at about 65 tok/s on an M3 Max 64 GB, perplexity within 0.94% of the 8-bit version, and 70/70 factual recall on the extraction test.
For actually deploying Hy-Memory-30B-A3B, the realistic bar is a 24 GB GPU or 32 GB Mac running Q4_K_M, while going to Q8_0 with long context requires 64 GB.
Hy-Memory-30B-A3B's 262K max context is a configured value; the GGUF uploader measured only up to 32K in practice.
Tencent's published comparison table shows Hy-Memory-30B-A3B scoring an average of 55.49 across six benchmarks, slightly above Claude Opus 5's 54.62, a gap of 0.87 points; but looking at individual items, 0.02 points under the Hy-Memory framework and 0.01 points under MemOS are essentially noise level.
Compared with Qwen3-30B-A3B, Hy-Memory-30B-A3B improves by an average of 11 points, and still scores 56.89 on the EverMemOS framework, which it never saw in training.
Community testing (34 English + 34 Russian conversations) shows that all quantized versions of Hy-Memory-30B-A3B can output valid JSON (68/68), but only 23/34 fully succeed.
In vLLM v0.31.0 testing, vllm serve tencent/Hy-Memory-30B-A3B --served-model-name hy-memory can start the BF16 checkpoint, with weights of about 60 GB, requiring one 80 GB card or two 48 GB cards with tensor parallelism.
Tencent open-sourced the Hy-Memory-30B-A3B model, a 30B-parameter mixture-of-experts (MoE) model that activates about 3B parameters per token.
Hy-Memory-30B-A3B itself does not store memories; instead it acts as a memory writer, specifically screening information to write into an external memory store and updating old entries.
Hy-Memory-30B-A3B is released under the Apache-2.0 license, requiring no extra legal communication for commercial use.
7 more timeline events
All Facts (17)
On the MLX side, upstream mlx-lm does not yet recognize hy_v3, but hermitdave's Mixed4-6 build bundles hy_v3.py and points to it via config, so it can be loaded directly with stock mlx-lm; the uploader reports decoding at about 65 tok/s on an M3 Max 64 GB, perplexity within 0.94% of the 8-bit version, and 70/70 factual recall on the extraction test.
50%UnverifiedFor actually deploying Hy-Memory-30B-A3B, the realistic bar is a 24 GB GPU or 32 GB Mac running Q4_K_M, while going to Q8_0 with long context requires 64 GB.
50%UnverifiedHy-Memory-30B-A3B's 262K max context is a configured value; the GGUF uploader measured only up to 32K in practice.
50%UnverifiedTencent's published comparison table shows Hy-Memory-30B-A3B scoring an average of 55.49 across six benchmarks, slightly above Claude Opus 5's 54.62, a gap of 0.87 points; but looking at individual items, 0.02 points under the Hy-Memory framework and 0.01 points under MemOS are essentially noise level.
50%UnverifiedCompared with Qwen3-30B-A3B, Hy-Memory-30B-A3B improves by an average of 11 points, and still scores 56.89 on the EverMemOS framework, which it never saw in training.
50%UnverifiedCommunity testing (34 English + 34 Russian conversations) shows that all quantized versions of Hy-Memory-30B-A3B can output valid JSON (68/68), but only 23/34 fully succeed.
50%UnverifiedIn vLLM v0.31.0 testing, vllm serve tencent/Hy-Memory-30B-A3B --served-model-name hy-memory can start the BF16 checkpoint, with weights of about 60 GB, requiring one 80 GB card or two 48 GB cards with tensor parallelism.
50%UnverifiedTencent open-sourced the Hy-Memory-30B-A3B model, a 30B-parameter mixture-of-experts (MoE) model that activates about 3B parameters per token.
50%UnverifiedHy-Memory-30B-A3B itself does not store memories; instead it acts as a memory writer, specifically screening information to write into an external memory store and updating old entries.
50%UnverifiedHy-Memory-30B-A3B is released under the Apache-2.0 license, requiring no extra legal communication for commercial use.
50%UnverifiedHy-Memory-30B-A3B uses the HYV3ForCausalLM architecture, with 128 routed experts and 8 activated per token, 48 layers, hidden dimension 2048, max position embeddings 262,144, and support for Chinese and English.
50%UnverifiedHy-Memory-30B-A3B is a task-specific model trained specifically as a Memory Provider, not a general-purpose conversational model.
50%UnverifiedHy-Memory-30B-A3B excels at memory management tasks such as generating conversation summaries, performing context compaction, and identifying and writing entities.
50%UnverifiedIn the Pi Blackhole system, Hy-Memory-30B-A3B can serve as a reflector or dropper, letting the main model focus on conversation generation.
50%UnverifiedFor consumer-grade graphics card users, loading the 30B weights is the biggest threshold, and VRAM usage issues may affect adoption.
50%UnverifiedIf Tencent releases a smaller-scale version (such as 7B or smaller) or a quantization-friendly deployment solution, the model's actual adoption rate may be higher.
50%UnverifiedTencent released the Hy-Memory-30B-A3B model on Hugging Face.
50%