Reading MiniMax M3.1 Flash's official benchmarks: mostly flagship rivals, mixed vs DeepSeek Flash
Only the two DeepSeek V4.1 Flash versions are Flash models among six rivals; MiniMax wins 2 of 6 against the newest one but tops Automation-Bench (62.2) and 256K long context (60.3).
Remember "Space Bunny", the mystery model that topped the leaderboards? It was an early version of MiniMax M3.1 Flash. Now we have the official benchmarks. Here's how to read them. It's not a pure Flash vs Flash fight. Only 2 of the 6 rivals are Flash models. Both are DeepSeek V4.1 Flash (the original and the new 1017 version). The rest are full-size models: • Kimi K3 and GLM-5.3 (big open-weight Chinese models) • GPT-5.6 Sol and Opus 5 (closed frontier models) So a small fast model is being compared with the top models in the world. That's bold. The catch: The newest DeepSeek Flash (1017) only has scores on 6 benchmarks. On those 6, MiniMax wins 2 and loses 4. Against the older DeepSeek Flash, MiniMax wins almost every row. Guess which one makes the chart look better. Where it really shines: • Automation-Bench: 62.2, best in the whole table • LOCA-Bench 256K (long context): 60.3, also the best • BankerToolBench: 76.2, tied with Opus 5 Opus 5 still tops most rows, especially coding and general reasoning. What the benchmarks test: • CoWork: agent and office-style tasks • Coding: real repo and software tasks • Cyber: security tasks • General: vision and hard reasoning Open or closed? Open. Weights drop on Oct 23. So the fair fight is really with DeepSeek Flash, which is also open. Against that, it trades blows. Against closed giants like Opus 5, it's punching way above its size. Price: • $0.15 per 1M input tokens • $0.60 per 1M output tokens • $0.015 per 1M cached tokens MiniMax says it's 1/3 the cost of M3 and 3x faster. For agent work that burns millions of tokens, that matters more than 2 points on a benchmark. My take: An open Flash model sitting this close to Opus 5 on agent tasks, at this price, is the real story. But these are the lab's own numbers. Wait for independent tests before you switch.
This post quotes another post: x.com/minimax_ai/status/2112150867868541221