MiniMax M3 shows open-source models catching frontier labs on agentic tasks

MiniMax released M3, its first multimodal open-source model trained from scratch with text and vision data, achieving natural cross-modal attention. Together AI, handling the majority of M3 inference traffic, reveals the optimization challenges behind serving models with sparse attention, 1M context windows, and agentic workloads.

The partnership highlights a growing pattern: model creators focus on post-training while inference providers handle deployment complexity. Together AI begins kernel optimization before launch, targeting KV cache management, attention kernels, and quantization. The shift from chat to agentic workloads with large codebase prompts changes inference priorities significantly.

MiniMax trained M3 for long-horizon tasks like replicating ICLR papers over 12-hour runs, using RL with carefully designed environments and reward functions. They monitor for intermediate progress and prevent hacking. The model's self-evolution capability allows internal use to accelerate development.

The panel argues open-source models like M3 can catch frontier labs, with Together AI's Dan predicting better GPU utilization and deeper inference optimizations ahead. However, concrete benchmarks on real-world agent effectiveness remain limited, and the 12-hour task replication claims lack independent verification.