MiniMax M3: Unverified Claims & China's Intelligence Law

MiniMax has launched its M3 foundation model, claiming it's the first open-weight system to combine frontier coding-agent performance, a one-million-token context window, and native multimodal capabilities. The API is live today. Here's the thing: developers need to consider several points. Benchmark scores are reported by MiniMax itself and run on their own infrastructure. The promised open weights have not yet been released. Also, China's National Intelligence Law requires MiniMax to cooperate with government intelligence work for all prompts processed through its API. The M3's core technology is MiniMax Sparse Attention, or MSA, which aims to make large context windows viable. MiniMax claims MSA significantly cuts per-token compute and accelerates processing and response generation. These speed claims, however, come from MiniMax's own benchmarking. The company has not yet published the details needed for outside engineers to verify these numbers independently. MiniMax reports M3 scored 59.0% on the SWE-Bench Pro benchmark, compared to 58.6% for GPT-5.5 and 54.2% for Gemini 3.1 Pro. These performance claims are also company assertions until independent verification is possible. The bottom line: developers should be aware of these factors before integrating M3 into their workflows.