A product manager describes a website in a paragraph and hits enter. Minutes later a working prototype is live — search bar, pages, and a deployable build — assembled by a single model that reasoned, wrote the code, and wired up the tools.
The lab behind that demo calls the approach "agent-native," and the model is GLM-4.5.
One model for reasoning, coding, and tool use
GLM-4.5 is a large model (大模型) from Zhipu AI (智谱华章), the Beijing company spun out of Tsinghua University, now using the international brand Z.ai. It launched on 28 July 2025 and the weights were open-sourced under the MIT license.
The design goal was unusual: instead of shipping separate models for math, coding, and tool use, Zhipu folded all three into one set of weights. The idea is that agentic applications — software that plans, calls functions, browses, and writes code over many turns — should not have to stitch together several specialised models.
Two sizes shipped:
- GLM-4.5: 355 billion total parameters, 32 billion active, mixture-of-experts (混合专家, MoE).
- GLM-4.5-Air: 106 billion total, 12 billion active — a lighter sibling for cheaper deployment.
Zhipu frames the efficiency directly: GLM-4.5 has about half the parameters of DeepSeek-R1 and a third of Kimi K2, yet it reports competitive results on several standard benchmarks.
Hybrid reasoning you can toggle
A distinctive feature is dual-mode (hybrid) reasoning. A caller can switch between:
- a thinking mode for hard reasoning and tool use, where the model deliberates before answering;
- a non-thinking mode for instant responses.
That matters for agents, because the same model can do a quick lookup and a deep multi-step plan without swapping models. The context window is 128K tokens, with up to 96K output tokens.
Native function calling is built in, and Zhipu says GLM-4.5 reached a 90.6% tool-calling success rate on its internal agent evaluations — the kind of number that matters more for real agent loops than for static Q&A leaderboards.
How it scored
On the 12-benchmark composite Z.ai published — covering agentic, reasoning, and coding tasks — the company reported GLM-4.5 ranked third overall and first among open-source models, behind only closed frontier systems. On individual tests Z.ai cited:
- MMLU-Pro: 84.6% — academic knowledge.
- SWE-bench Verified: 64.2% — real-world software engineering.
- τ-bench: ~70% — agent tool use (Retail 79.7, Airline 60.4).
- BrowseComp: 26.4% — open-ended web browsing, ahead of Claude 4 Opus (18.8%) in Z.ai's run.
The model was also pre-trained on roughly 23 trillion tokens — about 15T of general text followed by 7–8T of code and reasoning data — then sharpened with reinforcement learning.
Access and price
GLM-4.5 is available three ways:
- Chat: chat.z.ai and the Z.ai web app.
- API: through BigModel.cn (Z.ai's platform), with one-click compatibility claimed for the Claude Code framework.
- Weights: open on Hugging Face (zai-org/GLM-4.5) and ModelScope under MIT.
The published API price was about US$0.60 per million input tokens and US$2.20 per million output tokens for GLM-4.5, with the Air variant at roughly US$0.20 / US$1.10 — several times cheaper than comparable Western frontier tiers and free for experimentation on the Flash variants.
Why the MIT license matters
MIT is about as permissive as open model licenses get: self-host, fine-tune, and ship commercially with no royalty. For enterprises that cannot send data to a closed API, that changes the calculus. A hospital, bank, or government office can run GLM-4.5 on its own hardware and keep the data inside its walls — something a closed model cannot offer.
Honest limitations
- Benchmarks are vendor-reported. The third-overall and tool-use figures come from Z.ai's own runs; independent reproductions are thinner than for some Western models.
- 128K context, not 1M. Rivals like MiniMax-M1 and Gemini 2.5 Pro handle far longer single-pass context, which matters for whole-codebase or whole-document tasks.
- It has been superseded internally. Z.ai has since shipped GLM-4.6, GLM-4.7, and the GLM-5 line; GLM-4.5 is no longer the lab's flagship, though the weights remain live and MIT-licensed.
- Big to self-host. At 355B parameters, full-precision serving needs serious multi-GPU hardware; GLM-4.5-Air exists precisely to ease that.
- bilingual strength is the real edge. GLM-4.5 is genuinely strong on Chinese-English tasks, which is where it tends to beat similarly sized Western-open models most clearly.
What readers can do now
- Use the API or the free Flash tier first. Point a Claude-Code-compatible agent at BigModel.cn or chat.z.ai to feel the tool-calling and coding behaviour before committing to infrastructure.
- Self-host the Air variant if hardware is tight. GLM-4.5-Air (106B/12B) keeps most capability at a fraction of the memory and is the pragmatic on-prem choice for a single rich-GPU server.
- Keep data on your own metal with MIT. If compliance forbids sending prompts to a third-party API, pull the MIT-licensed weights and serve them behind vLLM or SGLang — no usage restrictions, no per-call royalty.