What open-weight models can match Opus 4.x or GPT 5.4 performance at a reasonable throughput?