Kimi Drops K2.7, Then MiniMax Unleashes M3 — Open-Source Code Models Are Insane

Yesterday Moonshot open-sourced Kimi K2.7-Code. Today it's MiniMax's turn. M3 is also open-weight, also a code model — but its approach is different. Instead of claiming 'best in class,' it says 'I want it all': coding + 1M context + native multimodal, all packed into one open-source model. First of its kind in the open-source camp.

💡 What You Will Learn

Yesterday Moonshot open-sourced Kimi K2.7-Code. Today it's MiniMax's turn. M3 is also open-weight, also a code model — but its approach is different. Instead of claiming 'best in class,' it says 'I wa

📜 Table of Contents

• M3's First Killer Feature: Coding That Goes Toe-to-Toe with Closed-Source Flagships
• Second Killer Feature: 1M Context, But the Architecture Is What Matters
• Third Killer Feature: Native Multimodal, Not a Bolt-On Patch

M3's First Killer Feature: Coding That Goes Toe-to-Toe with Closed-Source Flagships

M3: 428B total params, 23B active, MoE architecture. Key benchmarks: - SWE-Bench Pro: 59.0% — surpasses GPT-5.5 and Gemini 3.1 Pro, approaches Claude Opus 4.7 - Terminal Bench 2.1: 66.0% - MCP Atlas: 74.2% - KernelBench Hard: 28.8% - SWE-fficiency: 34.8% SWE-Bench Pro is the hardest real-world software engineering benchmark — it's not LeetCode, it's fixing real GitHub issues. M3 has crossed the GPT-5.5 line here.

Second Killer Feature: 1M Context, But the Architecture Is What Matters

1M token context isn't new (M2.7 had it). What's new is how M3 does it. M3 uses a self-developed MiniMax Sparse Attention architecture (MSA). Traditional full attention is O(n²) — more tokens = exponentially more compute. MSA's insight: not every token needs to 'look at' every other token. Most attention computation is wasted. Result: at 1M token context, M3's per-token compute is ~1/20th of the previous generation. Real engineering moat. For developers: a 500K-token codebase that would OOM or burn through compute on other models can run smoothly on M3 at an order of magnitude lower cost.

Third Killer Feature: Native Multimodal, Not a Bolt-On Patch

Many models add multimodal as an afterthought. M3 trained on text, images, and video from day one — expanded data scale and training pipeline. Supports image understanding, video understanding, and Computer Use — letting AI operate a desktop like a human, switching between apps to complete tasks. Combined with the first two, M3 covers three scenarios: 1. Coding + Agent (SWE-Bench Pro, MCP Atlas scores) 2. Long documents + large codebases (1M context + MSA compute advantage) 3. Multimodal + desktop control (native multimodal + Computer Use) Each of these alone has open-source contenders. But packaged in one open-weight model? M3 is the first.