100,000 Huawei Ascend 910B processors. Zero Nvidia GPUs. That's what Zhipu AI used to train GLM-5, a 744-billion-parameter language model that now sits among the top open-source models in the world.
Released on February 11, 2026, GLM-5 represents the clearest signal yet that US export controls on advanced chips have not stopped China from building frontier AI. The model was trained entirely on domestically produced hardware - Huawei chips designed by its HiSilicon subsidiary and manufactured by SMIC, China's largest chipmaker, using a 7-nanometer process.
The Benchmark Numbers
GLM-5 posts competitive scores across major evaluations:
- SWE-bench Verified (real-world software engineering): 77.8%
- AIME 2026 (advanced math competition): 92.7%
- GPQA-Diamond (graduate-level science questions): 86.0%
- Humanity's Last Exam (a notoriously difficult cross-domain test): 50.4% with tools
It also leads open-source models on BrowseComp, Vending Bench 2, and MCP-Atlas. These numbers put it in the same conversation as GPT-5.2 and Claude Opus 4.6 on several tasks - not matching them across the board, but close enough to matter.
Price Changes the Equation
At $1.00 per million input tokens and $3.20 per million output tokens, GLM-5 API access costs a fraction of what Western frontier models charge. For businesses processing large volumes of text - legal document review, customer support, content moderation - that pricing gap is significant. The model ships under an MIT license, meaning anyone can download and run it locally for free.
The market noticed. Zhipu AI's stock on the Hong Kong Stock Exchange jumped 28.7% within 24 hours of the release.
Beyond Language Models
GLM-5 is part of a broader pattern. China's AI progress in early 2026 extends across multiple domains. Baidu's Apollo Go robotaxi service has now completed 20 million rides across 22 cities, logging 190 million fully driverless kilometers. The service handles over 250,000 fully autonomous rides per week - comparable to Waymo's US numbers - and has reached per-vehicle profitability in Wuhan, where it operates over 1,000 vehicles.
The safety record is notable: Apollo Go's driverless vehicles average 10.14 million kilometers between airbag deployments, exceeding both human driver baselines and Waymo's published safety metrics.
What makes this moment different from previous rounds of "China is catching up" headlines is the self-sufficiency angle. GLM-5 proves a frontier-class model can be trained without any access to Nvidia hardware. The Huawei Ascend 910B chips are not as efficient as Nvidia's H100 or B200 - training likely took longer and cost more per unit of compute - but the end result is competitive. Export controls raised the difficulty level; they did not set a ceiling.
For AI tool users and businesses evaluating model options, the practical takeaway is straightforward: the pool of capable models keeps growing, pricing pressure is intensifying from both open-source releases and Chinese competitors, and the assumption that frontier AI requires Western hardware no longer holds.