Zhipu Releases GLM-5: Open-Source 744B Model

Zhipu AI, a leading Chinese artificial intelligence laboratory, released GLM-5, a massive open-source language model that represents a watershed moment in China's AI development. The model features 744 billion parameters in a mixture-of-experts (MoE) architecture with 44 billion active parameters, a context window of 200K tokens, and was trained entirely on Huawei Ascend 910B processors — without any NVIDIA chips. Released under the MIT license, GLM-5 claims performance parity with Claude Opus 4.5 and GPT-5.2 on coding and agent tasks.

Model Architecture and Performance

GLM-5 employs a mixture-of-experts design that activates only 44 billion of its 744 billion total parameters per inference, balancing massive knowledge capacity with computational efficiency. The architecture includes a 200K token context window for processing long documents and codebases, and achieves 77.8% on SWE-bench Verified, an industry-standard benchmark for coding ability.

Zhipu claims GLM-5 surpassed Google DeepMind's Gemini 3 Pro on multiple internal evaluation metrics, with particular strength in coding, agent tasks, and long-context understanding. While independent third-party validation would strengthen these claims, the benchmarks demonstrate that Chinese AI labs are matching or approaching Western frontier model performance.

Training on Huawei Ascend: No NVIDIA Required

The most geopolitically significant aspect of GLM-5 is its training infrastructure. Zhipu trained the entire 744B parameter model on Huawei Ascend 910B processors, proving that frontier AI development can proceed without NVIDIA GPUs — the dominant hardware for AI training worldwide.

This carries profound implications for the effectiveness of U.S. export controls on advanced AI chips. Washington has restricted the sale of high-end NVIDIA processors to China, aiming to slow Chinese AI development. GLM-5's existence demonstrates that domestic semiconductor alternatives like Huawei's Ascend platform can support frontier model training at scale, potentially undermining the strategic rationale behind chip export restrictions.

Open Source Under MIT License

Zhipu released GLM-5 under the MIT license, making it fully open source for academic and commercial use worldwide. This decision contrasts sharply with the proprietary approaches of OpenAI, Google, and Anthropic. The open-source strategy positions Zhipu as a democratizer of frontier AI, offering developers globally — especially in regions where proprietary Western models face restrictions — a powerful alternative built on non-sanctioned hardware.

The MIT license means developers can freely use, modify, and redistribute GLM-5, lowering barriers to entry for AI development and enabling international collaboration on model improvements.

Market Reaction

Investors responded enthusiastically to the GLM-5 announcement: Zhipu's Hong Kong-listed shares surged 28.7% to HK$402, reflecting confidence in the company's technical progress and competitive positioning. The market reaction validates the strategic bet on open-source frontier models as a viable business approach.

Geopolitical Implications

GLM-5's release underscores several geopolitical realities. U.S. chip export restrictions have not prevented China from developing frontier AI. Domestic semiconductor alternatives to NVIDIA are viable at the frontier scale. Multiple independent AI ecosystems are emerging globally rather than a single Western-dominated paradigm. The AI technology race between the U.S. and China continues to intensify, with both nations producing competitive frontier models.

Policymakers and strategists will scrutinize GLM-5 as evidence of how the global AI landscape is fragmenting into parallel development ecosystems, each with distinct hardware, software, and governance approaches.

What This Means for AI Engineers

For AI professionals, GLM-5 signals that model diversity is increasing rapidly. Frontier capabilities are no longer monopolized by a handful of Western companies, and open-source models are competitive with proprietary alternatives. Engineers who understand diverse model architectures, training methodologies, and hardware platforms will have a significant advantage.

If you're preparing for roles in AI development, InterviewAlly helps you practice the kind of technical depth across AI systems and architectures that distinguishes expert engineers in an increasingly competitive and global field.