What Happened
MiniMax launched M3 through its API and developer products on June 1, 2026, according to its English announcement. The post stated that the technical report and downloadable weights would follow. This entry records the model and API launch, rather than assuming that weights were already public that day.
Why It Matters
M3 combined long context, native visual input and agentic software work in a new architectural generation. Its significance lies in making context scaling part of the attention design, while extending the model beyond text-only coding to workflows involving images, video and desktop interaction.
Technical Details
MiniMax Sparse Attention, or MSA, selects blocks of key-value information instead of applying full attention across the entire context. The model supports up to one million tokens and can switch thinking on or off. The release targets coding, research and office tasks; reported efficiency gains depend on the benchmark and implementation.