Model Overview & Capabilities
MiniMax-M3 is a frontier artificial intelligence model engineered by MiniMax, officially verified in March 2026 and evaluated on the Artificial Analysis Intelligence Index v4.3 with an overall score of 30.0. Architected as a proprietary API service, the system operates across a context window of 1000k tokens, allowing enterprise developer agents and autonomous code evaluation tools to ingest comprehensive repository structures, cross-language modules, and multi-layered documentation hierarchies without degradation of semantic coherence.
Under standardized software engineering benchmarks—including SWE-bench Verified, Terminal-Bench 2.0, Aider Polyglot, and LiveCodeBench—MiniMax-M3 exhibits rigorous multi-step problem solving, deterministic SEARCH/REPLACE diff compliance, and robust tool-use navigation in sandboxed execution environments. Its reasoning capabilities are calibrated to minimize hallucination rates while handling complex syntax transformations, boundary conditions, and continuous integration workflows.
Economically and operationally, MiniMax-M3 generates output tokens at a median throughput of 186 tokens per second, with an observed initial latency of 1000 milliseconds. Artificial Analysis rates its estimated cost per task at $0.51 USD, backed by an API token pricing structure of $0.4 per million input tokens and $1.2 per million output tokens (resulting in an effective blended rate of $0.6 per million tokens).