MiniMax M3 slashes coding costs with 10x faster inference

MiniMax, a rising Chinese AI startup, unveiled its newest flagship model, M3, on June 1, 2026, targeting enterprise developers who build and maintain complex coding pipelines. M3 is architected from the ground up for long-context comprehension – handling entire codebases of over 100,000 tokens without losing coherence. The company claims a 10x reduction in inference cost per token compared to its previous generation, achieved through a novel sparse attention mechanism and a tiered memory system that caches frequently accessed code patterns.

Beyond raw efficiency, M3 is designed to automate multi-step development workflows. It can manage branching logic, refactor legacy code, and even detect regression bugs in real time. Shortly after launch, MiniMax announced a strategic partnership with Ant Group’s Alipay, integrating M3 directly into Alipay’s developer console. This allows Alipay’s 80,000+ partner developers to request code reviews, generate test suites, and debug production incidents using natural language, with M3 handling the heavy lifting within the payment ecosystem’s stringent latency requirements.