GLM-5
Open-weight frontier MoE model from Zhipu AI (清华系 Beijing-based AI lab). **744B total parameters, 40-44B active per inference token.** Trained on 28.5T tokens entirely on **Huawei Ascend chips** — the first frontier model built without any NVIDIA hardware in the training stack. Incorporates DeepSeek's Dynamically Sparse Attention (DSA) for efficient long-context handling up to 200K tokens. Maximum output length 131K tokens. MIT license, downloadable weights — Zhipu's bet runs counter to the typical Chinese AI-vendor "API-only" pattern.
Definition
Open-weight frontier MoE model from Zhipu AI (清华系 Beijing-based AI lab). **744B total parameters, 40-44B active per inference token.** Trained on 28.5T tokens entirely on **Huawei Ascend chips** — the first frontier model built without any NVIDIA hardware in the training stack. Incorporates DeepSeek's Dynamically Sparse Attention (DSA) for efficient long-context handling up to 200K tokens. Maximum output length 131K tokens. MIT license, downloadable weights — Zhipu's bet runs counter to the typical Chinese AI-vendor "API-only" pattern.
Two structurally important things at once. **First**, GLM-5 proves that a frontier-class model can be trained outside the NVIDIA ecosystem — using Huawei Ascend chips end-to-end — which de-risks the AI supply chain for China and reshapes the long-term sovereignty calculus for non-US AI infrastructure. **Second**, the MIT-licensed open-weight release is a strategic outlier among Chinese AI labs (most ship API-only): Zhipu chose to bid for global developer mindshare via weights rather than hosted-API revenue. The 2026 framing has crystallized this as China's open-weight flagship.
Agentic engineering workloads (the model is positioned explicitly as Zhipu's "Agentic Engineering" flagship), long-context document analysis up to 200K tokens via DSA, on-prem deployment on Huawei Ascend clusters in China, MIT-licensed alternative to DeepSeek-V3/V4 for derivative work, and any sovereign-cloud deployment where avoiding US-controlled silicon supply chain matters.
Recent developments
- Released February 11, 2026 — 744B total / 40-44B active MoE, 28.5T training tokens. Scaled from GLM-4.5's 355B/32B-active to 744B/40B-active. Per Lushbinary developer guide.
- First frontier model trained entirely on Huawei Ascend, no NVIDIA. End-to-end training on Huawei Ascend chips — establishes the precedent that frontier-class MoE training is feasible outside the NVIDIA hardware stack. Per Lets Data Science — China-trained frontier AI.
- SWE-Bench Verified 77.8% — outperforms many US-frontier models. GLM-5 outperforms Claude Opus 4.5 and Gemini 3 Pro on several benchmarks; HLE 30.5, AIME 2026-I 92.7, GPQA-Diamond 86.0. Per MorphLLM — GLM-5 2026.
- MIT license + downloadable weights — outlier among Chinese AI vendors. Most Chinese labs ship API-only; Zhipu's open-weight release drives adoption outside China and positions GLM-5 as a global open-weights option. Per Digital Applied — GLM-5 vs GPT-5.2/Claude 4.5.
- 200K-token context via DeepSeek DSA mechanism. Adopts the Dynamically Sparse Attention pattern from DeepSeek for efficient long-context handling without the dense-attention computational overhead. Per Apiyi — GLM-5 API guide.
Connections 4
Outbound 3
Inbound 1
enables1