deepseek-ai/deepseek-v4-pro

Full precision, max quality. KV cache is only ~10GB even at 1M tokens (sparse attention + Engram memory). Self-host economics close ~3–4B tokens/day — below that the API is cheaper.