On July 28, 2025, Zhipu AI (智谱) released GLM-4.5 and aimed at a specific complaint about large models (大模型): they were good at one thing each. One reasoned well, another coded well, a third called tools well, but none did all three in one set of weights. GLM-4.5 was built to close that gap — an "agent-native" model where reasoning, coding, and tool use share the same brain.
The model shipped open-weight under the MIT license, with a lighter sibling, GLM-4.5-Air, for cheaper deployment. Zhipu reported it as the top open-source model on a 12-benchmark aggregate at launch.
Key takeaways
- GLM-4.5 is a 355-billion-parameter MoE with 32 billion active per token; GLM-4.5-Air is 106B total / 12B active. Both are open weights (开源权重) under MIT.
- It is a hybrid-reasoning model with a thinking mode (complex reasoning, tool use) and a non-thinking mode (instant replies), built as an agent-native (智能体) model.
- Zhipu positions it as agent-native: one model unifies reasoning, coding, and tool use, with a reported 90.6% tool-calling success rate.
- Aggregate 12-benchmark score was reported third overall and first among open models; API pricing started at about US$0.11 (HK$0.88) per million input tokens.
One model, three jobs
GLM-4.5 was trained in stages on roughly 23 trillion tokens in total: 15 trillion tokens of general pre-training, then about 8 trillion tokens of code, reasoning, and agent data, followed by reinforcement learning to sharpen those skills. The design goal was explicit — satisfy the "fast-rising agentic applications" that need a model to plan, write code, and call tools inside one request.
Zhipu notes that GLM-4.5's 355B parameters are about half of DeepSeek-R1's and one-third of Kimi K2's, arguing that higher parameter efficiency, not raw size, drove the results. The smaller GLM-4.5-Air (106B / 12B active) was reported to reach levels comparable to Gemini 2.5 Flash and Claude 4 Sonnet on reasoning benchmarks.
Hybrid reasoning, again
Like Qwen3, GLM-4.5 offers two modes. Thinking mode handles hard reasoning and tool use; non-thinking mode gives fast direct answers. Callers pick per request. Context length is 128K tokens, with up to 96K output tokens.
For agents, the model adds native function calling and one-click compatibility with the Claude Code framework through Zhipu's BigModel.cn platform — a practical nod to developers who already build agent workflows.
The agentic benchmarks
Zhipu measured agent ability on τ-bench (TAU-bench) and BFCL-v3 (Berkeley Function Calling Leaderboard v3), reporting GLM-4.5 matched Claude 4 Sonnet on both. On BrowseComp, a web-browsing benchmark, GLM-4.5 answered 26.4% correctly with tool access — above Claude-4-Opus (18.8%) and close to o4-mini-high (28.3%).
Reported scores include:
- MMLU-Pro: 84.6%.
- AIME 2024: 91%.
- SWE-bench Verified: 64.2%.
- TAU-bench: 79.7% (Retail), 60.4% (Airline).
- BrowseComp: 26.4%.
- Tool-calling success rate: 90.6%.
On a 12-benchmark aggregate (agentic, reasoning, coding), Zhipu reported GLM-4.5 third overall and first among open models, with GLM-4.5-Air sixth.
Cost and access
The API launched at roughly ¥0.8 per million input tokens and ¥2 per million output tokens — about US$0.11 (HK$0.88) input and US$0.28 (HK$2.2) output. Weights are on Hugging Face and ModelScope. The combination of MIT licensing and a sub-355B active footprint makes self-hosting realistic for teams that cannot run a full 1T-class model.
Why it matters
GLM-4.5 is part of a 2025 pattern: Chinese labs racing not just on size but on "agent-native" design — models that treat tool use as core, not bolted on. By unifying three skills and open-sourcing the weights, Zhipu gave smaller teams a single downloadable model for agent builds that previously needed several specialized ones.
The efficiency claim is the part worth watching. Zhipu argues that folding reasoning, coding, and tool use into one 355B model beats running three separate specialist models on both cost and coherence — a single context means the model remembers the code it wrote while it calls the API. If that holds in production, it lowers the barrier for startups building autonomous workflows without a frontier-model budget. The Air variant extends the same idea downward: a 106B model that a mid-size cluster can serve, narrowing the gap between open and closed on reasoning while staying self-hostable.
Honest limitations
- BrowseComp at 26.4% is still low; web-browsing agents from this model are far from reliable on hard queries.
- The benchmark aggregate and most scores are Zhipu's own, under its harness and attempt counts; independent runs may differ.
- Like its peers, GLM-4.5 is text-only; no native vision or audio at launch.
- The 355B full model still needs serious hardware; the Air variant trades some peak skill for lighter deployment.
What readers can do now
- Pull GLM-4.5-Instruct (or GLM-4.5-Air if hardware-limited) from Hugging Face and test one agent task that needs both code and tool calls in a single prompt.
- Compare the MIT-licensed open weights against a closed agent model on your real tool-calling set before committing to a paid API.
- If you already use Claude Code-style workflows, try Zhipu's one-click compatibility to see if an open model meets your agent build.