MiniMax

MiniMax is the AI model company hosting the Developer Meetup captured in 对话 MiniMax 闫俊杰：M3、10X 计划、10T 模型、和智能的终局. Yan Junjie, its founder and CEO, describes the company through model iteration from M1 to M2/M2.7 and MiniMax M3, with AI coding, editing, token usage, and real engineering feedback serving as practical signals for model usefulness.

当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent adds Adao and Zeying as MiniMax practitioners discussing MiniMax Agent, Agent Harness, Multi-Agent Collaboration, and Interleaved Thinking alongside Hermes Agent. The source frames MiniMax as wanting its models to work well inside external agent frameworks such as Hermes Agent, Open Cloud, and Open Claw, not only inside its own agent product.

当可靠的代码变成了偶尔发疯的OpenClaw，我们未来的工作范式变迁 adds MiniMax as a low-friction deployment example. A host uses a MiniMax deployment surface to publish an AI-generated HTML tool, but the platform’s safety checks modify the script and introduce a first-run bug, illustrating that easy AI deployment still needs debugging and clear platform responsibility boundaries.

136. 全球大模型季报第9集：和广密聊，Coding是AGI第二幕、硅谷御三家真相、模型正成为新一代OS adds MiniMax to the domestic-model convergence story. The source says MiniMax, Kimi, Zhipu AI, and other Chinese model companies are increasingly moving toward the Anthropic-style high-value-task and agent/coding route rather than only broad consumer chat.

Source Position

• MiniMax is presented as trying to make developer-facing meetups a continuing activity rather than a one-off event.
• Its M2/M2.7 work is described as focused on coding and editing rather than balanced coverage of every general chat scenario.
• MiniMax M3 is treated as an ambitious next step whose usage is ahead of expectation while capability is still approaching the desired state.
• The company frames models and harnesses/agents as mutually reinforcing paths toward higher intelligence.
• In safety, finance, law, and other domains, MiniMax is expected to bring more domain experts into model development through Domain Expert Alignment.
• The Hermes Agent discussion says MiniMax M2.5/M2.7-class agentic capability helped the domestic OpenCloud/OpenClaw wave feel newly useful.
• MiniMax is described as using model plus harness in model-development work, supporting the wiki’s Agent Self-Evolution and Model Harness Co-Evolution themes.
• The company is also described as using internal digital employees to find open-source projects that can connect to MiniMax models.
• The Keji Luandun OpenClaw episode uses MiniMax less as a model-company case and more as an example of low-threshold AI deployment where platform safety layers can change generated code.
• Episode 136 treats MiniMax as part of a domestic model-company shift toward coding, agents, and high-value work.

Connections

• Yan Junjie — founder and CEO speaking for the company in the roundtable.
• MiniMax M3 — model discussed through coding use, cost, and capability tradeoffs.
• Model Harness Co-Evolution — MiniMax view that model and agent/harness progress reinforce each other.
• Frontier Model Scaling — scaling discussion around 3T and 10T training.
• AI Coding Verification — engineering bottleneck identified as AI coding adoption grows.
• AI Interpretability By AI — long-term intelligence and safety question raised by Yan.
• Adao and Zeying — MiniMax speakers in the Hermes Agent roundtable.
• Hermes Agent, Open Cloud, and Open Claw — external agent contexts MiniMax wants its models to support.
• Multi-Agent Collaboration, Interleaved Thinking, and Agent Self-Evolution — themes added by the Hermes Agent source.
• AI Coding Verification and Probabilistic Software — deployment and platform-safety boundary added by Keji Luandun.
• Kimi, Zhipu AI, Doubao, AGI Three Acts, and Model As Operating System — domestic-model strategy frame added by episode 136.