MLX: Opus 4.5 implementing DeepSeek-V3.2, here follow the Opus comments and video of first run on M3 Ultra 512GB! 🔥 - Conversion: ✅ Successful (4.502 bits/weight, 88 shards) - Generation: ✅ Working at ~21-23 tokens/sec - Peak memory: ~379 GB (this is a 671B parameter model!)