chapter seventeen

17 The Open Model Revolution of 2026

 

This chapter covers

  • GLM-5: a 744B-parameter open model trained entirely on Chinese chips
  • Qwen3-Coder-Next: an ultra-sparse MoE model for coding agents
  • MiniMax M2.5: SWE-bench parity with proprietary models at 1/20th the cost
  • The economics of open models and what they mean for AI agents
  • The Unsloth-to-Ollama pipeline for fine-tuning and deploying custom models
  • Practical steps for trying these models yourself

In February 2026, the AI landscape shifted fundamentally. Within a single ten-day window, multiple open-source models reached -- and in some cases surpassed -- the performance of the best proprietary AI systems. For anyone who has been building with local models, this chapter explains what happened, why it matters, and what you can do with these new capabilities.

17.1 February 2026: A Turning Point

Until early 2026, there was a widely held assumption in the AI industry: proprietary models from companies like OpenAI, Anthropic, and Google would always be ahead of open-source alternatives. Open models were useful for learning and experimentation, but they were not frontier. Not state-of-the-art.

That assumption collapsed in February 2026.

17.2 GLM-5

17.2.1 Architecture

17.2.2 Trained on Huawei Chips

17.2.3 Benchmark Results

17.2.4 License

17.3 Qwen3-Coder-Next

17.3.1 Ultra-Sparse MoE Architecture

17.3.2 Coding Agent Focus

17.4 MiniMax M2.5

17.4.1 SWE-bench Parity

17.4.2 Cost Advantage

17.5 The Benchmark Picture

17.6 The Economics of Open Models

17.6.1 Agent Cost Analysis

17.6.2 Why Local Agents Change What Is Possible

17.7 The Unsloth-to-Ollama Pipeline

17.7.1 The Complete Custom Model Workflow

17.7.2 Why Customization Matters Now

17.8 What This Means for You

17.8.1 How to Try These Models

17.8.2 Practical Next Steps

17.9 Further Learning

17.10 Beyond Individual Models: The Rise of Orchestration

17.11 Summary

17.12 Exercises