MiniMax-M3

MiniMax-M3 will be released in June 2026 and is an architecture-level generational jump after the M2 series. M3 adopts MiniMax's self-developed sparse attention architecture MSA (MiniMax Sparse Attention), which expands the context window to 1 million tokens, while reducing the calculation amount of a single token to about 1/20 of the previous generation model. The pre-filling stage is accelerated by more than 9 times, and the decoding stage is accelerated by more than 15 times. M3 is the first open source model in China that has the three core capabilities of "cutting-edge coding capabilities, 1M ultra-long context, and native multi-modality". It supports image and video input, and can operate desktop computers. On programming benchmarks, M3 achieved SWE-Bench Pro 59.0% and Terminal-Bench 2.1 66.0%, and showed preliminary usability in office workflow and financial fields.

Context

1000000 tokens

Max output

0 tokens

Input

llm

Output

llm