# GLM-5 — Frontier Open Weights Trained on Ascend, Not NVIDIA

Times of AI Desk · 2026-02-12 · Models

[https://timesof.ai/2026/02/zhipu-ai-releases-glm-5-745b-open-frontier-model](https://timesof.ai/2026/02/zhipu-ai-releases-glm-5-745b-open-frontier-model)

> Zhipu AI (Z.ai) released GLM-5 (~745B open-weight MoE, Feb 11–12): trained entirely on Huawei Ascend + MindSpore—no NVIDIA. Competitive coding/agentic claims vs proprietary peers; MIT-style open licensing reported; HK-listed shares jumped ~28–34% on the news. Sovereignty stack as the trade.

Export controls were supposed to freeze China’s frontier stack at the GPU layer. Zhipu’s proprietary demonstration with GLM-5: **frontier-class open weights trained entirely on domestic Ascend**—capability without the NVIDIA bill of materials.

**Zhipu AI (Z.ai)** released **GLM-5** around February 11–12. Reported scale: ~**744–754B** parameters (MoE in some descriptions)—commonly rounded ~**745B**. Trained exclusively on **Huawei Ascend** with **MindSpore**—company and wire framing as self-reliant stack under U.S. export controls. Positioning: strong coding and long-running agent tasks; tops several open-source leaderboards in contemporaneous reports; approaches or matches aspects of leading proprietary models at substantially lower cost (vendor/coverage). Open licensing (MIT-style references in coverage); weights via Hugging Face and Chinese platforms. Market: Hong Kong-listed shares rose sharply (**~28–34%** gains cited across reports); demand reportedly overwhelmed initial capacity, with pricing adjustments in some channels.

## Claims vs checks

Ascend/MindSpore training claim and open release are **company + Reuters-linked primary reporting**. Parameter counts vary slightly across reports (~744–754B)—cite ranges, don’t invent a single audited figure. Bench “matches Claude/GPT/Gemini” claims are **vendor/secondary**—verify on Artificial Analysis / OpenRouter rankings for the February window. Stock moves are market fact; they are not model quality.

## Limits

- Domestic-chip training success ≠ identical serving economics outside China.
- Open license terms: confirm exact SPDX on the model card, not headlines.
- Evaluation transparency and safety practice details remain thinner than U.S. system cards.

## Sources

- Company and Reuters-linked reporting on February 11, 2026 GLM-5 release (Zhipu / Z.ai): parameter count, Ascend + MindSpore, benchmark claims.
- Coverage of stock reaction and demand surge (28–34% gains cited).
- Hugging Face and OpenRouter references confirming availability and relative ranking in February 2026.