Hy-Memory-30B-A3B

Released 2026-10-15 on Hugging Face (API createdAt 2026-10-15T09:52:00Z; card last modified 2026-10-16). Affordable-inference class: 30B MoE named A3B, same shape as the 30B-A3B reference class, and the card states ~3B active parameters. Not a general chat model — it is a specialized long-term-memory writer that replaces the general-purpose LLM on the write path of Mem0, MemOS, and Hy-Memory (Chat, Working, Coding) and transfers to held-out EverMemOS. Trained in two stages: Memory On-Policy Distillation, then Attributed Memory Reinforcement Learning (arXiv 2610.22196, submitted 2026-10-15). config.json (model_type hy_v3, HYV3ForCausalLM): hidden 2048, 48 layers, heads 32 / kv heads 4 / head_dim 128 so gqa_ratio 8, intermediate_size 6912 on the dense MLP, moe_intermediate_size 1536, expert_hidden_dim 768, 128 routed experts, 8 per token, 1 shared expert, first_k_dense_replace 1 (mlp_layer_types = 1 dense + 47 sparse), vocab 120832, max_position_embeddings 262144, rope_theta 11158840 with rope_type default and no rope_scaling (native), hidden_act silu, tie_word_embeddings false, qk_norm true. Safetensors total 30.065B, recorded as params_b 30.06. active_params_b 3.0 is the card's '~3B active' figure, not a count derived from the config — the shared-expert width is not given as a parameter count, so treat 3.0 as the vendor's rounded active figure. KV (bf16, every layer is attention): 48 layers x 4 kv heads x 128 head_dim x 2 x 2 bytes = 98304 B/token = 96 KiB/token. At 262144 context that is ~24.6 GiB of KV before quantization, so the long context is the expensive part, not the 3B active weights. flops_per_token_active = 2 x 3.0e9 = 6.0e9 is the schema approximation off the vendor's ~3B figure. THROUGHPUT: the card gives no tok/s. BENCHMARKS (vendor_claim, fixed 20% subsample, six-metric harness mean): macro average 55.49 versus Claude Opus 5 at 54.62, DeepSeek V4.1 Flash 53.64, Qwen3.6-35B-A3B 50.37, Qwen3-30B-A3B 44.37. Apache-2.0.