The Image generated by AIYXL using AI
MiniMax released M3.1 Flash on October 19, focusing the update on applications such as multi-step task execution, software development, and computer operation.
According to the official announcement, the model is already available through the MiniMax API, M Plan, and MiniMax Code, and the weights will be released later. Therefore, this launch means users can use the model through the relevant services; it does not mean the weights can already be downloaded for self-deployment.
From content generation to continuous execution
The tests disclosed by MiniMax cover cross-software business processes, code modifications, and computer operation, emphasizing the model's ability to maintain task consistency across multiple rounds of tool calls.
These scenarios impose requirements on the model that differ from single-turn Q&A: understanding the goal is only the beginning; the system also needs to process tool-returned results, connect operation steps, and check the final outcome.
However, the benchmark results in the announcement come from vendor disclosures and cannot be directly equated with the model's performance in all enterprise environments.
The image is sourced from the MiniMax official website
After the price decline, the cost of completing tasks still needs to be calculated
MiniMax says the overall usage cost of M3.1 Flash for agent tasks is 70% lower than the previous generation, and online inference speed has also improved. The above comparison uses the vendor's test conditions; actual costs will still vary with task length, number of retries, and tool-calling methods.
AIYXL observation: Cost competition among agent models is extending from how much a single call costs to how many resources are needed to complete a piece of work.
For users, a report that can be adopted directly or a code modification that passes checks is more meaningful than simply generating more text. During evaluation, model fees, execution time, failed retries, and manual modifications can be recorded together to see whether the final outcome meets requirements.
This launch of M3.1 Flash provides a new testing option; its actual value still needs to be verified through specific work.
Source: https://www.minimax.io/blog/