MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input
What changed
MiniMax launched M3.1 Flash, a massive mixture-of-experts (MoE) model with 428 billion parameters total. Despite that scale, it activates only around 23 billion parameters per input. The model supports an unusually large 1 million token context window, extending practical use cases that need long-term memory or large documents. API pricing is set at $0.15 per million input tokens and $0.60 per million output tokens. Model weights will be released on October 23.
Why builders should care
M3.1 Flash offers one of the largest MoE architectures available with agentic capabilities, meaning it can orchestrate complex multi-step reasoning or autonomous actions within applications. The sparse activation allows handling massive parameter counts without the latency and cost penalties typical of dense models this size. One million-token context is a rare feature that benefits workloads involving entire books, extensive code bases, or multi-document synthesis. At $0.15 per million tokens input, it’s priced competitively, putting pressure on the cost of high-context, large-model access.
The practical takeaway
Businesses with large context requirements for summarization, research, or agentic automation may find M3.1 Flash worth testing. Its sparse active parameters reduce compute needs on each query while keeping model quality high. The large context window opens new opportunities for applications that traditionally struggled with token limits. The API pricing signals affordability but output costs remain higher, so practical deployments will need to balance input depth against response length.
What to watch next
Watch for user feedback on the trade-offs between activated parameter sparsity and real-world performance in tasks demanding long-range context. The October 23 weight release allows deeper inspection and potential fine-tuning, which may accelerate adoption by research and enterprise teams. Competitors offering dense models with shorter contexts or other MoE models will likely adjust pricing or capabilities to challenge MiniMax here. Developers should track integrations, benchmarks, and the ecosystem forming around ultra-long context transformer use.
AI Quick Briefs Editorial Desk