For years the unwritten rule of large models was specialization: one model reasoned well, another coded well, a third ran agents well, and stitching them together was the engineer's job. Zhipu (智谱) bet against that rule with GLM-4.5, a single open-weight model built to do all three at once — and to do it cheaply enough that teams actually deploy it.
The wager is that the next wave of value is not a smarter chatbot but a model that can plan, call tools, and write working code inside one coherent reasoning loop.
Key takeaways
- GLM-4.5 is a mixture-of-experts (混合专家, MoE) model with 355B total parameters and 32B active; a lighter GLM-4.5-Air carries 106B total / 12B active.
- Both are released under the MIT License as open weights on Hugging Face and ModelScope.
- The model is a hybrid-reasoning (混合推理) system with a thinking mode for complex reasoning and tool use and a non-thinking mode for instant replies.
- It offers a 128K context window and native function calling, and Zhipu reports agentic scores matching Claude 4 Sonnet on key benchmarks.
One model, three jobs
GLM-4.5's headline claim is unification. Rather than shipping separate reasoning, coding, and agent models, Zhipu trained a single base that scores competitively across all three. On a battery of a dozen benchmarks spanning agentic, reasoning, and coding tasks, the company places GLM-4.5 third overall among models it compared — behind only the largest closed systems — and first among open-weight models.
The architecture is MoE, which is why a 355B-parameter model can be served with only 32B active per token. That gap is the whole economic story: you get frontier-class capacity without paying to activate all of it on every request. GLM-4.5-Air pushes the idea further, hitting reasoning levels the company compares to Gemini 2.5 Flash and Claude 4 Sonnet at roughly a third of the flagship's total parameter count.
Agentic by design
Where GLM-4.5 is most interesting is agentic work. It ships with a 128K context and native function calling, and Zhipu points to two benchmarks — τ-bench (TAU-bench) and BFCL-v3, the Berkeley Function Calling Leaderboard — where it says GLM-4.5 matches Claude 4 Sonnet. On BrowseComp, a hard web-browsing benchmark, Zhipu reports GLM-4.5 answering 26.4% of questions correctly with tool access, above Claude 4 Opus at 18.8% and close to o4-mini-high at 28.3%.
Those are vendor-reported numbers, but they describe a consistent shape: a model tuned to plan across many turns and use tools mid-reasoning, then summarize. That is exactly the behavior agent builders complain is missing from general-purpose chat models.
The MIT license advantage
GLM-4.5 and GLM-4.5-Air are both released under the MIT License — about as permissive as open-weight licensing gets, with none of the usage caps that some custom licenses attach above certain thresholds. Weights are on Hugging Face and ModelScope, and the API is available through Zhipu's BigModel.cn (大模型开放平台) platform, which the company says is one-click compatible with the Claude Code framework for agent workflows.
Pricing, where stated, is low: Zhipu listed API rates of about 0.8 yuan per million input tokens and 2 yuan per million output tokens — roughly US$0.11 (HK$0.9) in and US$0.28 (HK$2.2) out — with a high-speed tier reaching up to 100 tokens per second. Those figures make GLM-4.5 a plausible default for cost-sensitive agent deployments, though exact rates shift over time and by region.
Why it matters outside China
Zhipu frames GLM-4.5 as a foundation model for the fast-rising agentic application layer, and the open-weight release means teams anywhere can download it, fine-tune it, and run it inside their own infrastructure. For organizations with data-residency or budget constraints, a permissively licensed model that performs near strong closed systems removes a common excuse for staying locked to a single vendor's API.
What readers can do now
- Evaluate GLM-4.5 on multi-turn agent tasks, not single questions — its native function calling and 128K context are where it pulls ahead of general chat models.
- Compare the Air variant first if cost or latency is tight; 106B/12B active can match much larger models on reasoning at a fraction of the bill.
- Check the MIT terms before shipping, confirm current API pricing on BigModel.cn, and self-host the open weights where data must stay in-house.
Honest limitations
This article draws on Zhipu's official GLM-4.5 blog (z.ai), the National Center for Science and Technology Infrastructure notice, and Shanghai Securities / Beijing News reporting that reprinted the launch facts. The parameter counts (355B/32B and 106B/12B), the MIT license, the 128K context, the 26.4% BrowseComp figure, and the 0.8/2 yuan per-million-token pricing are reported as the vendor states them and were not independently re-measured. Benchmark comparisons to Claude 4 Sonnet, Claude 4 Opus, Gemini 2.5 Flash, and o4-mini-high are drawn from Zhipu's own charts and should be read as vendor-reported. I have not run GLM-4.5 myself, have not verified the "third overall / first open-weight" ranking on a current leaderboard, and have not covered GLM-4.5's later revisions. Treat the numbers as a starting point and run your own eval on the agentic tasks you care about.