Cohere Embed Multilingual v3.0

Config.json is minimal: only hidden_dim 1024 and n_positions 512. Multilingual (100+ languages). Supports cosine, dot product, and Euclidean similarity. Not open weights — available via Cohere API, AWS Bedrock, and Azure AI. Same output dimension and context limit as the English variant. No GGUF or ONNX ports (closed model). kv_cache_bytes_per_token = 0 for the same reason as cohere-embed-english-v3 - single-pass embedding, no decode loop, no per-token cache. Same caveat: config.json publishes only hidden_dim 1024 and n_positions 512, so this is a task-based inference rather than a config-derived figure. The two Cohere v3 records share output_dim and context length; only the training languages differ.