Nomic Embed Text v2 MoE

HEAD/KV FIELDS FILLED, with the reason they were null made explicit: config declares no num_key_value_heads and no head_dim, because causal FALSE and use_cache aside this is a bidirectional encoder that scores documents independently - there is no incremental decode and no KV cache to size. num_kv_heads 12 and gqa_ratio 1.0 are recorded from n_head 12 (MHA, nothing to compress) so the field set is complete, but they are NOT inputs to kv_cache_bytes_per_token, which stays 0.0 for this record. head_dim 64 (768/12) was already filled on the same basis. shared_expert = false is now POSITIVE rather than inferred: config has num_shared_experts 0, an explicit zero. MoE embedding model: 8 experts, top_k 2, moe_every_n_layers 2 (so 6 of 12 layers are MoE). Total params 494M but active params only ~58M per token — this is the key serving advantage. At 58M active params, flops_per_token_active is ~1.16e8, comparable to bge-small (33M) but with the capacity of a much larger model. Matryoshka-trained: supports dimension truncation. Vocab 250048 (multilingual). Apache-2.0. GGUF and ONNX ports exist. CONTEXT_SCALING_METHOD (native): config carries no rope declaration of any kind for this encoder.