Cohere Embed English v3.0

Config.json is minimal: only hidden_dim 1024 and n_positions 512. No architecture details, no parameter count, no layer count. English-only. Supports cosine, dot product, and Euclidean similarity. Not open weights — available via Cohere API, AWS Bedrock, and Azure AI. The 1024-dim output and 512-token limit are the serving-relevant facts. No GGUF or ONNX ports (closed model). kv_cache_bytes_per_token = 0 is recorded because this is an embedding model: a single forward pass produces one vector, there is no autoregressive decode and no per-token cache. Recorded explicitly rather than left null so a reader querying for cache-free models finds it, but note the config is minimal (only hidden_dim 1024 and n_positions 512) so the 0 rests on the embedding-task classification, not on a read architecture - the encoder layer count is genuinely unknown from config.