tencent/Hy-Memory-30B-A3B model card and config.json

Primary source for the Hy-Memory-30B-A3B record. HF API createdAt 2026-10-15T09:52:00Z, lastModified 2026-10-16, pipeline_tag text-generation, cardData license apache-2.0, safetensors total 30,064,725,888 parameters (BF16 30,064,719,872 + F32 6,016). Model card (fetched 2026-10-16) calls it a compact long-term-memory writer with ~3B active parameters, trained to replace the general-purpose LLM on the write path of Mem0, MemOS, and Hy-Memory, and transferring to held-out EverMemOS. config.json (resolve/main, same fetch) confirms model_type hy_v3, architectures HYV3ForCausalLM, hidden_size 2048, num_hidden_layers 48, num_attention_heads 32, num_key_value_heads 4, head_dim 128, intermediate_size 6912, moe_intermediate_size 1536, expert_hidden_dim 768, num_experts 128, num_experts_per_tok 8, num_shared_experts 1, first_k_dense_replace 1, max_position_embeddings 262144, vocab_size 120832, hidden_act silu, rope_theta 11158840, rope_type default, tie_word_embeddings false. mlp_layer_types is 1 dense then 47 sparse. Benchmark table on the card (55.49 macro average vs 54.62 for Claude Opus 5 on a fixed 20% subsample) is a vendor claim. Technical report is arXiv 2610.22196, submitted 2026-10-15.