Zhipu debuts open-source model GLM-5 in race with Gemini, Claude

China’s Zhipu AI has launched its latest AI model, GLM-5, amid a race among Chinese tech firms before the Spring Festival.

The company claims the model shifts AI development from “vibe coding” to “agentic engineering,” with improved coding and automation performance.

According to Zhipu, GLM-5 outperformed Google DeepMind’s Gemini 3 Pro in internal tests but lagged behind Anthropic’s Claude on coding benchmarks.

The model’s size doubled from 355 billion to 744 billion parameters and was trained on 28.5 trillion tokens using a new DeepSeek Sparse Attention architecture.

GLM-5 is open-sourced on GitHub and Hugging Face and can be deployed on Chinese semiconductor chips, including Huawei and Baidu’s Kunlunxin.

The launch follows Zhipu’s recent Hong Kong IPO, valuing the company at about US$18 billion.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

This model is built for practical work, not just conversation

Agent Mode supports “agentic engineering” and creates office files like .docx, .pdf, and .xlsx directly from prompts 1.
Z.ai connects this utility focus to reinforcement learning (a method for training AI via feedback on actions). It says a new system called “slime” speeds up training for complex, agent-like behaviors 1.
On one benchmark, it hits a record-low hallucination rate by choosing to abstain instead of making up answers 1.
Pricing on OpenRouter (a marketplace for accessing multiple AI models via a single API) as of Feb. 11, 2026 runs about $0.80 to $1.00 per 1 million input tokens and $2.56 to $3.20 per 1 million output tokens. VentureBeat’s comparison puts that around five to six times cheaper on input than Anthropic’s Claude Opus 4.6 1.

A parallel, non-Western AI ecosystem is becoming viable

GLM-5 works with Chinese chips, which Z.ai frames as strategic amid U.S. export controls and China’s push for technological independence 2.
Zhipu AI says it trained GLM-Image fully on a domestic stack using Huawei’s Ascend Atlas 800T A2 devices (AI accelerator servers) plus the MindSpore AI framework (Huawei’s machine-learning software toolkit) 3.
These moves could split the global tech supply chain, which can push business and IT leaders to weigh geopolitical risk when choosing between separate AI ecosystems 1.
That divide could also steer different AI priorities, with Chinese firms leaning toward scalable execution and automation while Western labs lean toward cognitive depth and reasoning 1.