MiniMax M2 - Efficient Frontier Models

Time: 12:20 PM

Speaker Bio: Represents MiniMax, a Chinese AI company.

Speaker Profile: Full Speaker Profile

Company: MiniMax develops efficient AI models. M2 is claimed to run 2x faster than Claude Sonnet at 8% of the cost. Recently achieved top agentic evaluation scores.

Focus: New frontier models emphasizing efficiency. M2 is particularly focused on agentic workloads.

References:

• YouTube Link
• MiniMax Website

Slides

Slide: 12-30

Key Point: MiniMax’s M2 model can scale across multiple specialized agents (research, web development, reporting) working collaboratively, while remaining cost-effective for long-running agentic tasks.

Literal Content:

• Title: “M2 Scales and Collaborate in Multi-Agent Systems”
• Three columns showing different agent types:
  • “Research Agent” (left): Screenshot of a research interface with conversation and data
  • “Web Development Agent” (center): Dashboard showing “Run dashboard” with sections for Research, Web Development Agent, and Report Agent
  • “Report Agent” (right): Screenshot of a detailed report document
• Bottom text: “Small & Cost-effective to run long agentic tasks”
• QR code in center labeled “MiniMax Agent”
• MiniMax logo and “Intelligence with Everyone” tagline

Notes

• First Chinese model
  • Many modalities, and also applications, 150M users
• Research and developers sitting side by side working on things
• Lot of post training on MiniMax
• How does this square with prompts across harnesses
• Training shapes M2 behavior
  • Coding experience
    • Scaled environments and scaled experts
    • Pulled down the data from GitHub/internet
    • Scaled expert development as a feedback training
  • Interleaving thinking
    • In the real world the environment is often noisy and dynamic
    • Multiple rounds of tool called and reacts to the environment
  • Agent Generalization
    • Tool scaling?
    • Doesn’t scale if we change a tool scaffold
    • Adapt to perturbations across the entire model operational space
    • Apply perturbation pipelines
  • Multi-agent scalability
    • Small and cost effective
• Better models for the community to use