GitHub radar
MiniMax M3.1 Flash: a Chinese model for agents and cybersecurity
MiniMax released a new flagship model for coding, agent work, computer use, and cybersecurity, and compared it directly to Opus 5 from Anthropic — its own benchmark table tells a mixed story.
MiniMax, a Chinese AI lab, released M3.1-Flash, a large mixture-of-experts model the company pitches for coding, agent work, computer use, and cybersecurity — four jobs in one release. The model has about 428 billion parameters in total, though only 23 billion work on any single token at a time. It reads text, images, and video, and holds up to 1 million tokens of context. MiniMax's own benchmark table compares it to its Chinese rivals DeepSeek, Kimi, and GLM, plus GPT-5.6 from OpenAI and Opus 5 from Anthropic (the company behind Claude). The results are uneven: M3.1-Flash beats Opus 5 on an automation test, 62.8 against 50.3, and ties it exactly on a banking-tools test, 76.2 to 76.2. But on the two abilities MiniMax highlights as its strengths — computer use and cybersecurity — the model trails: nine points behind Opus 5 on computer-use tasks, and behind a rival Chinese model, GLM-5.3, on the cybersecurity test.
Why a vibe-coder should care
I work in Claude Code on Opus for daily tasks and Fable, Anthropic's hardest-task model, for the toughest jobs, so a Chinese model's own benchmark table matters more to me than its marketing copy. This one is honest in an odd way: MiniMax wins a couple of narrow tests and loses the two it built the release around. That pattern repeats with almost every Chinese model release I look at — loud framing at the top of the page, then a table that goes both ways underneath. And since the model only makes sense on datacenter hardware, trying it means opening MiniMax's own web agent, not running anything yourself.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open MiniMax M3.1 Flash's web agent (https://agent.minimax.io) and give it a real coding or file-handling task, the kind I'd normally hand to an agent in Claude Code — show me the result.
This is a server-only model with about 428 billion parameters — you cannot run it at home, it needs data-center-level hardware.
Open on Hugging Face▌ More finds