MiniMax M3 Drops: Beats GPT-5.5 on Coding, Million-Token Context, But Users Aren't Buying It

M3 scored 59.0% on SWE-Bench Pro, surpassing GPT-5.5 and Gemini 3.1 Pro, and even beat Claude Opus 4.7 on SVG-Bench. MiniMax had M3 independently reproduce an ICLR 2025 award paper — it ran autonomously for 12 hours, producing 18 commits and 23 experiment charts. The context window reaches 1 million tokens thanks to the in-house MSA sparse attention architecture, which reduces per-token computation to 1/20th of the previous generation while speeding up prefilling by 9x and decoding by 15x. MiniMax also gave M3 a broken Triton kernel to optimize — M3 worked for 24 hours, submitted 147 benchmarks and 1959 tool calls, boosting Hopper FP8 hardware utilization from 7.6% to 71.3% with zero human intervention. Despite the impressive tech, the community response is surprisingly divided. Two main complaints: First, benchmarks don't equal real-world experience. Many developers say they no longer trust benchmarks. Media tests found M3's generated racing game looked okay but was barely playable. Multimodal vision is detailed but lags behind DeepSeek and Qwen3.7 on specific visual QA tasks. Second, the price hike. M3's API is priced at two tiers: 4.2 RMB per million input tokens for contexts under 512K, doubling to 8.4 RMB for 512K-1M. Token plans were also adjusted, with users reporting faster token consumption. In 2025, MiniMax reported $79.04 million in revenue (up 159% YoY), but losses surged 302% to $1.87 billion — the brutal reality of the LLM race where revenue doubles but losses double too. MiniMax recently filed for A-share IPO guidance in Shanghai, aiming to become the first listed LLM company on China's A-share market.

M3's technical foundation is solid — MSA sparse attention, native multimodal, and autonomous agent training are genuine architectural breakthroughs. But MiniMax's core challenge is commercial, not technical. The M2 series won developer mindshare with low prices, but M3 raised them. That's a dangerous pivot — squeezed between OpenAI/Anthropic/Google ahead and DeepSeek/Qwen open-source behind, raising prices without top-tier real-world experience is a risky bet.