MiniMax on Margrop Blog

MiniMax M3 Officially Released: Demystifying the MSA Sparse Attention Architecture, Plus a Look Inside the Mavis Sandbox

Mon, 01 Jun 2026 22:00:00 +0800

On June 1st, MiniMax (Xiyu Technology) officially released its next-generation general-purpose model, MiniMax M3. It simultaneously maxes out three traditionally hard pillars — frontier coding, ultra-long context, and native multimodality — and it is the only fully open-source model in the world to do so. On the same day, I spent a few hours poking around inside Mavis, mapped out the invisible sandbox behind it, and bundled the experience together with my recent usage data and an invite link.

Disclaimer: This article is 100% written and generated by MiniMax-M3 (including all code blocks, technical diagrams, and structure), without any manual edits or changes.

Update (2026-06-01 22:00): A few hours after this post went live, the team published a follow-up announcement acknowledging that the original Token Plan migration was communicated poorly and that weekly-quota handling for long-time users wasn’t done right. They’ve laid out compensation, more quota, and a refund channel. I’ve added Section 9 — “Token Plan Revamp: Latest Announcement” at the bottom of this post. If you’re on Token Plan, or about to subscribe, read it before you spend another yuan.

Double-Consumed Streaming: Debugging a NewAPI v1.0.0-rc.10 MiniMax Proxy Bug

Sun, 31 May 2026 18:00:00 +0800

TL;DR

An AI Agent was producing duplicated replies in DingTalk — not two separate messages, but a single message whose content appeared twice. The root cause traced back to a streaming bug in NewAPI (the QuantumNous fork of OneAPI) v1.0.0-rc.10: when proxying MiniMax through the OpenAI Chat Completions protocol, the finish chunk contained both delta.content and message.content with identical values. The Agent’s stream handler consumed both as “visible text,” concatenating the content twice. The fix was straightforward: switch the OpenClaw provider protocol from openai-completions to anthropic-messages, which OneAPI natively supports and where streaming behaves correctly.

No internal IPs, tokens, model IDs, or private paths appear in this article. All configuration snippets have been sanitized.

Why My AI Assistant Repeated One Reply Four Times: Debugging an OpenClaw WeChat Channel Issue

Fri, 29 May 2026 11:24:29 +0800

Short version

The symptom looked like a messaging-channel bug: an AI assistant connected to a personal chat channel appeared to repeat the same reply multiple times. The real bug was earlier in the chain. Before the reply ever reached the message-sending layer, the OpenClaw agent had already produced duplicated visible text. The useful fix was not to patch the chat sender blindly, but to split the path into transport, session, model-routing, and provider layers, then compare the OpenAI-compatible gateway path with the native provider path. Once the faulty compatible provider candidate was removed from the visible model set and the personal IM channel was pinned to the native provider, the same minimal prompt returned exactly once.

This post is deliberately privacy-safe. It contains no real internal addresses, account IDs, tokens, session IDs, personal chat identifiers, or private file paths. Configuration examples use placeholders. The value is the debugging method, not the private environment.

One Month with OpenClaw on MiniMax: Stable, Fast Enough, and Valuable Because It Lasts

Thu, 19 Mar 2026 10:30:00 +0800

Over the past month, I have kept MiniMax as the primary model behind OpenClaw and used it across a full set of automation tasks: routine inspections, connectivity troubleshooting, health checks, security hardening, scheduled-task cleanup, and even blog writing support. Looking back, the biggest gain was not “how much work AI can replace,” but a more practical conclusion: for a system like OpenClaw that needs to run continuously, cooperate with tools, and handle a large amount of Chinese context, the most important thing is not raw benchmark strength. It is the balance between stability, response speed, integration cost, and long-term maintainability.