MiniMax M2.7 and Zhipu's GLM-5: The Agent-First Chinese Open Models

DeepSeek and Qwen get the headlines, but China's open-source AI depth goes further. Two names every builder should know in 2026: MiniMax and Zhipu (now branded Z.ai).

MiniMax: Compact and Agentic

MiniMax M2 is a compact mixture-of-experts model aimed at coding and agentic workflows, open-sourced on Hugging Face at aggressive pricing (~$0.30 / 1M input, ~$1.20 / 1M output). The newer MiniMax M2.7 (open-sourced April 12, 2026) reports 56.22% on SWE-Pro and 57.0% on Terminal Bench 2 — strong agentic-coding numbers for a small, cheap, open model.

Zhipu / Z.ai: The GLM Line

Zhipu's GLM family has moved fast: GLM-4.6 (357B params, MIT license, 200K context) → GLM-4.7 (December 2025) → GLM-5 (February 2026). The MIT licensing is notable — it's about as permissive as open weights get, which matters for commercial deployment.

The IPO Subplot

Both Zhipu and MiniMax timed model releases around Hong Kong listing plans. That's a sign of the times: Chinese AI labs are maturing from research shops into public companies, and shipping competitive open models is part of the investor pitch.

The theme across MiniMax and Zhipu: small, cheap, agent-tuned, and open. These aren't 2-trillion-parameter behemoths — they're models you can actually run and afford.

Why 'Agent-First' Matters

The 2026 frontier isn't just about answering questions — it's about doing tasks: reading a codebase, running commands, iterating on failures. Models like M2.7 and GLM-5 are explicitly tuned for that loop, which makes them practical engines for automation even without frontier-tier raw intelligence.

The Bottom Line

If you're building coding agents or workflow automation and want open weights you can host cheaply, MiniMax and Zhipu belong on your shortlist next to Qwen and DeepSeek.

Brandomize builds agentic automation for content and operations. The best model for an agent is often a small, cheap, tuned one — not the biggest. We'll help you find the fit.