China's Zhipu AI to step up AGI push in 2026

Zhipu AI said it will increase efforts toward artificial general intelligence (AGI) in 2026 and continue open sourcing its AI models after its planned IPO in Hong Kong.

The announcement was made during a Reddit AMA, where researchers said they would keep releasing model weights and sharing technical results with the open-source community.

The Beijing-based startup recently launched GLM-4.7, the latest version of its flagship model, which it said matched Anthropic’s Claude Opus 4 on the SWE-Bench benchmark.

Zhipu AI said it is dedicating more computing resources to pre-training, and added that the next major update will be called GLM-5 only if improvements are significant.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Zhipu’s open-source release leaves open the compute gap with US labs

Zhipu says GLM-4.7 ties Claude Opus 4 on SWE-Bench, yet R&D was 1.59 billion yuan in H1 2025 1.
It spent over 1.1 billion yuan on cloud fees 2. The outlay means reliance on rented gear over owned GPU clusters (large groups of graphics processors used to train and run AI models).
Zhipu plans more pre-training (the compute-intensive initial training phase for large models), yet export limits on advanced chips could cap access, and with no hardware details the 2026 AGI push is hard to judge.

Cloud providers outside China can host GLM-4.7 with compliance guarantees

GLM-4.7 weights are on Hugging Face (a widely used repository for AI models) and ModelScope (an open model hub) 3. vLLM (a high-throughput inference engine), SGLang (a scalable serving stack for large language models), and Ollama (a lightweight runtime for running models locally or on servers) 4 run it.
In a llama.cpp (a C++ project for running large language models efficiently) GGUF (a compact, quantized model format) setup, it needs 130GB VRAM for full GPU offloading (keeping the entire model resident in GPU memory) 5.
Zhipu serves 12,000 enterprise customers and 45 million developers 1. MLOps (machine learning operations) platforms can sell managed fine-tuning (task-specific adaptation), guardrails (safety plus policy enforcement), and multilingual support to buyers seeking alternatives to direct API (application programming interface) access from Chinese providers.