OpenAI's gpt-oss family (20B / 120B) with MXFP4 MoE, attention-sinks, built-in tools via Responses API