MiniMax M2 Interleaved Thinking

Interleaved MiniMax Demo

Interactive demo showcasing MiniMax-M2's interleaved thinking for agentic workflows, demonstrating adaptive strategy after each tool call with transparent traces and cost benchmarking. The project distills what we learned while building context-aware agents for front-end teams: keep the loop observable, keep the tooling grounded, and quantify efficiency against other LLMs.

---

Why this exists

• Explain interleaved thinking in practice. Every reasoning burst, tool call, and result is streamed to the terminal and logged to disk, so practitioners can see why MiniMax-M2 course-corrects faster than linear agents.
• Demonstrate agent-native workflows. MiniMax calls bespoke tools (design tokens, component specs, pattern guidance) to build a design brief, exercising the same MCP/shell/browsers-style chains we see in production.
• Benchmark against other coding LLMs. The run summary emits live token counts and the equivalent MiniMax pricing (0.3 $/MTok in, 1.2 $/MTok out) so you can compare against GLM 4.6, K2 Thinking, Claude Sonnet 4.5, etc.
• Serve as a starter kit for OSS contributions. The code is intentionally small, well-documented, and easy to extend with additional tools or scenarios.

---

Key learnings encoded here

1. Interleaved > linear: forcing the model to think after every tool result drastically reduces redundant calls in long-horizon front-end tasks.
2. Grounded tools beat fabricated answers: all tools read from claude_minimax/examples/sample_project/, ensuring explanations are backed by source material.
3. Observability builds trust: color-coded CLI output + JSONL logs make it trivial to review or share how MiniMax-M2 solved a task.
4. Cost transparency matters: developers need concrete $/token math when deciding between C