all-MiniLM-L6-v2

THE DEFAULT EMBEDDER: 6 layers, 384 hidden, 22.7M params — the smallest model in this set. kv_cache_bytes_per_token = 0, KV formula does not apply. 384-dim output = 1.5 KB per document. This is the model everyone starts with because it is fast, small, and good enough for many use cases. The 6-layer depth means it is 2x faster than bge-small (12 layers) at the cost of some retrieval quality. Apache-2.0. GGUF and ONNX ports exist. The sentence-transformers library default. kv_cache_bytes_per_token = 0 is a VERIFIED value, not an unfilled one: this is a bidirectional BERT encoder (6 layers, config num_attention_heads 12), so there is no causal decode loop and therefore no per-token KV cache to plan for. The KV formula does not apply because there is no generation step - the whole document is encoded in one pass and the cost is a single attention over n_positions, not a cache that grows with it. Contrast a 6-layer causal decoder of similar size, which would carry 2 x 6 x 12 x 32 x 2 = 9,216 B/token.