Zhipu AI Releases GLM-5: A 744B Open-Source Model That Rivals Western AI Giants

Chinese AI lab Zhipu AI has released GLM-5, a 744-billion-parameter large language model available under the permissive MIT license. The model, released on February 12, 2026, was trained entirely on Huawei Ascend chips and claims near-parity with leading Western models from Anthropic, OpenAI, and Google on key benchmarks — a milestone that signals China's rapidly closing gap in frontier AI development.

Key Highlights

• 744 billion total parameters with 40 billion active at inference using Mixture-of-Experts (MoE) architecture
• MIT license enabling free commercial use, fine-tuning, and community development
• Trained entirely on Huawei Ascend chips, achieving full independence from US-manufactured semiconductor hardware
• 77.8% on SWE-bench Verified, approaching Claude Opus 4.5's 80.9%
• Record-low hallucination rate on the Artificial Analysis Intelligence Index v4.0

Technical Specifications

GLM-5 nearly doubles the parameter count of its predecessor GLM-4.5, which had 355 billion parameters. The model was trained on 28.5 trillion tokens, up from 23 trillion in the previous generation. It features a 200,000-token context window and a 131,000-token output capacity, making it one of the most capable open-source models for handling long documents.

The model uses Deepseek Sparse Attention (DSA) to reduce deployment costs while maintaining performance across long-context tasks. It runs on both NVIDIA GPUs and Chinese-made chips from Huawei, Moore Threads, and Cambricon.

Benchmark Performance

GLM-5 positions itself as a serious contender against the best proprietary models:

The model outperforms both Deepseek-V3.2 and Kimi K2.5 on several coding and agent task benchmarks. Notably, Zhipu claims GLM-5 leads the entire industry — including US competitors — on knowledge reliability, achieving a -1 score on the AA-Omniscience Index, a 35-point improvement over its predecessor.

Five Core Capabilities

GLM-5 excels in five areas that Zhipu identifies as critical for enterprise adoption:

1. Creative writing with stylistic versatility across formats
2. Advanced code generation and debugging
3. Multi-step reasoning for complex problem solving
4. Agentic intelligence with autonomous planning and tool use
5. Long-context processing for extensive documents and research materials

The model also features a native "Agent Mode" that can convert raw prompts or source materials into professional office documents, generating ready-to-use .docx, .pdf, and .xlsx files directly.

Strategic Significance

The release carries weight beyond technical benchmarks. GLM-5 was trained using the MindSpore framework on Huawei's Ascend chips, demonstrating that China's domestic compute stack can produce frontier-scale models without relying on US semiconductor exports — a key concern since US export controls tightened in 2023.

The timeline is also notable: GLM-5 arrived roughly three months after the latest US flagship model releases, cutting the typical seven-month lag between Western and Chinese frontier models in half.

Availability and Pricing

Model weights are publicly accessible on Hugging Face and ModelScope under the MIT license. GLM-5 is also available on OpenRouter, priced at approximately $0.80–$1.00 per million input tokens and $2.56–$3.20 per million output tokens.

Developers can integrate GLM-5 with popular tools including Claude Code, OpenCode, and Roo Code through the OpenClaw framework. The official chat interface is available at Z.ai.

What's Next

The release of GLM-5 intensifies the global race in open-source AI. With Chinese firms like Zhipu, Deepseek, and MiniMax all releasing competitive models in rapid succession, the frontier of open-weight AI is no longer defined solely by Western labs. For developers and enterprises, this means more options, lower costs, and accelerating innovation in the open-source AI ecosystem.

Source: The Decoder