MiniMax M3 Medium reaches 73.17% F1 on DeepSearchQA, matching GPT-5 High

MiniMax M3 Medium's 73.17% F1 result represents a meaningful checkpoint in the evolution of open and commercial reasoning models. The score places it at parity with GPT-5 High Reasoning, which the DeepSearchQA paper reports at 73.24% F1. However, the body explicitly acknowledges that MiniMax M3 Medium lags behind the top-tier agents: Gemini Deep Research Agent (81.90% F1) and GPT-5 Pro High Reasoning (78.98% F1). The fully-incorrect rate of 19.01% also suggests room for improvement compared to those leading systems.

The evaluation methodology adds transparency to the claim. Rather than a paper assertion, the results are backed by 2,700 trials across 900 tasks, with all intermediate artifacts (trajectories, grades, prompts) published for inspection and reproduction. This open artifact approach—coupled with the You.com MCP tools and an optimized research skill tuned to the model and tool surface—establishes a clear foundation for further iteration. The total cost of $478.41 and median latency of 54.3 seconds provide operational context: the model operates within a reasonable resource envelope for research tasks, though the 95th percentile latency of 238.8 seconds indicates variability in harder queries.

The fact that the evaluation is described as competitive with GPT-5 High Reasoning, not with the top agents, sets realistic expectations. The body frames this as a step forward for the MiniMax M3 Medium line, not a breakthrough that overtakes the field. For practitioners, the key signals are the public artifacts (enabling validation and fine-tuning) and the documented cost and latency trade-offs.