Add q4-8emb pre-quantized variant (4-bit MoE + 8-bit embed, 870MB, parity 1.0)

Add bf16 weights (mirror of openai/privacy-filter)

Add model card, config, tokenizer (bf16 root)

initial commit