MiniMax M3: Frontier Coding, 1M Context, Native Multimodality

MiniMax released M3. The model reaches frontier performance on coding and agentic work. It uses MSA, a sparse attention architecture developed by the company, and supports context windows of up to 1 million tokens. M3 is natively multimodal. It accepts image and video input and can operate a desktop computer. The company states that M3 is the first and only open weight model to combine these three capabilities. M3 shows gains over the prior version in coding. It approaches leading closed source models on bug fixing, front end and back end development, and performance optimization. In agentic tasks, it handles common office workflows such as search and office suite operations and shows initial usability in finance. On benchmarks, M3 scored 59.0 percent on SWE Bench Pro, 66.0 percent on Terminal Bench 2.1, 34.8 percent on SWE fficiency, 28.8 percent on KernelBench Hard, and 74.2 percent on MCP Atlas. The company developed an interactive simulator to train and test models on multi turn collaboration patterns that mirror real developer workflows. MSA partitions key value blocks precisely and uses a KV outer gather Q operator. At 1 million tokens, per token compute falls to one twentieth that of the prior model. The company reports speedups exceeding 9 times in prefilling and 15 times in decoding while matching full attention on most capabilities. In internal tests, M3 reproduced an ICLR 2025 award winning paper over nearly 12 hours. It produced 18 commits and 23 experimental figures while matching reported trends and verifying a mitigation method. It also optimized an FP8 matrix multiplication kernel on Hopper GPUs over about 24 hours, completing 147 benchmark submissions and 1,959 tool calls from a basic starting point.