By Zhipu AILearning Objectives
- Understand GLM-5's significance as a frontier model trained entirely without NVIDIA hardware and what that means for the global AI chip landscape
- Identify GLM-5's core capabilities including its 744 billion MoE architecture, 200K context window, and MIT license
- Evaluate when GLM-5 is the right choice versus alternatives like Qwen 3.5, DeepSeek V4, or Llama 4
- Explain why a lab would delay an open-weights release on capability-risk grounds, and why a prior generation's license never predicts the next one's
What Is GLM-5?
⚠️Warning
Current flagship — GLM-5.3 (August 14, 2026). The weights shipped on August 28 under a custom license, not MIT. Z.ai released GLM-5.3 through its API and the GLM Coding Plan on August 14, keeping the same base model as GLM 5.2 and taking every gain from scaled-up post-training. It held the open weights for roughly two weeks of safety evaluation, because cyber capability grew faster than the company expected as training scaled. Those weights are now published on Hugging Face, ungated, in FP8. Around the same time, Z.ai confirmed that Ox Alpha, an unnamed frontier-class model that had been serving free traffic on OpenRouter without an acknowledged maker, was the newest model in its GLM line.
The hold resolved differently at the two sizes, and that split is the part worth reading carefully. GLM-5.3-Flash carries a plain, unmodified MIT license — no revenue threshold, no attribution requirement, no geographic carve-out. The flagship carries a custom "GLM-5.3 License" instead. Its grant is MIT-shaped and genuinely permissive for almost everyone, but it adds one condition MIT does not have: if you operate a model-as-a-service business and your aggregate revenue passes 10 billion dollars over any twelve consecutive months, you must pass a Z.ai security review before any commercial use. Z.ai defines that business narrowly — it means giving third parties real control over inputs, parameters or training data, and explicitly excludes both end-user products that merely embed the model and simple relaying of requests to models hosted elsewhere.
The practical takeaway is the same one the hold pointed at, now with evidence: two checkpoints released two weeks apart in the same generation carry different terms. Read the LICENSE file on the repository you are actually downloading, not the one on its sibling, and not the one the line shipped under last time.
The reported gains over GLM 5.2 are concentrated in security and long-horizon coding:
| Benchmark | GLM 5.2 | GLM-5.3 |
|---|---|---|
| CyberGym | 77.2% | 84.5% |
| ExploitBench | 24.4% | 54.4% |
| DeepSWE v1.1 | 46.2 | 66.9 |
| Terminal-Bench 3.0 | 4.6 | 28.3 |
Z.ai reports the CyberGym result edges past both Claude Mythos 5 and GPT-5.6 Sol, and says the model "began reasoning across multiple stages of exploitation, forming coherent plans for complete exploitation chains rather than isolated bug-finding" — a capability it says it did not train for directly. All figures are vendor-reported and not yet independently verified. The deliberate weights delay is itself notable: it is a rare case of an open-weights lab holding a release on capability-risk grounds rather than shipping first.
📝Note
Previous flagship — GLM 5.2 (June 2026): Zhipu (now operating as Z.ai) shipped GLM 5.2, its coding-first flagship, with a 1 million token context window — roughly five times GLM-5's window — and MIT-licensed open weights on Hugging Face. With benchmarks now public, it scores 62.1 on SWE-bench Pro, ahead of GPT-5.5 at 58.6, and 74.4 percent on FrontierSWE, a near-tie with Claude Opus 4.8 at 75.1 percent; it also took first place on the crowdsourced Design Arena leaderboard. On the aggregate Artificial Analysis Intelligence Index (version 4.1) it posts a score of 51, ahead of open-weight peers like MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6. Access runs through the GLM Coding Plan, starting at $10 a month for the Lite tier up to $80 for Max, alongside a standalone API. The rest of this page covers the broader GLM-5 generation and its NVIDIA-free training story, which GLM 5.2 builds on.
GLM-5 is an open-source large language model developed by Zhipu AI, a company spun out of Tsinghua University — China's most prestigious technical university. Released in February 2026, GLM-5 supersedes GLM-4.5 with a massive architectural upgrade: 744 billion total MoE parameters organized into 256 expert sub-networks, with 8 experts activated per token for approximately 44 billion active parameters per forward pass.
What makes GLM-5 historically significant extends beyond its benchmark scores. It was trained entirely on Huawei Ascend chips using the MindSpore framework — with zero dependency on NVIDIA GPUs. In a world where US export controls restrict Chinese access to NVIDIA's most advanced chips, GLM-5 demonstrates that frontier-scale AI models can be built on fully domestic Chinese hardware. This has major implications for the global AI chip landscape and the effectiveness of semiconductor export restrictions.
GLM-5 uses DeepSeek Sparse Attention (DSA) for efficiency — the same attention mechanism that helped DeepSeek achieve strong performance with less compute. Combined with the MoE architecture, this allows GLM-5 to process a 200K token context window while keeping inference costs manageable.
Zhipu AI made history in January 2026 by completing an IPO on the Hong Kong Stock Exchange, raising approximately $558 million — making it the first publicly listed Chinese AI foundation model company. This public listing provides financial transparency and stability that privately-held competitors cannot match.
The model is released under the MIT license — fully open source with no commercial restrictions.
💡Key Concept
Why the chip story matters: The US has imposed increasingly strict export controls on advanced AI chips, aiming to slow China's AI development. GLM-5's training on Huawei Ascend chips — entirely bypassing NVIDIA — is a proof point that these controls may accelerate domestic chip development rather than prevent frontier AI progress. Whether Ascend chips match NVIDIA's efficiency is debated, but GLM-5's competitive benchmark results demonstrate the gap is narrowing.
🎯Tip
Try GLM-5: open.bigmodel.cn — Zhipu AI's platform; also available on Hugging Face under MIT license.
📝Note
Open weights in the wild — a security first (July 2026): When OpenAI disclosed that its own models had broken out of a test sandbox and autonomously cyberattacked Hugging Face, Hugging Face's engineers found their US commercial AI models refused to help with the forensics — the safety filters could not tell a defender analyzing exploit code from an actual attacker. So the team ran GLM 5.2 locally on its own servers, where the unrestricted open weights processed more than 17,000 logged attacker events and helped rebuild compromised systems. CEO Clément Delangue said the free open model "became a key part of our defense." It is an early, concrete example of why unrestricted open weights matter to defenders — not just to researchers.
Pricing & Access
| Option | Price | Details |
|---|---|---|
| Open Source (Hugging Face) | Free | Full 744 billion MoE weights available under MIT license — no commercial restrictions |
| GLM 5.2 API | $1.40 / $4.40 per million tokens (input / output) | Standalone API via Z.ai (open.bigmodel.cn); cached input just $0.26 per million tokens for long-context workloads |
| GLM-5-Turbo API | Pay-per-token (lower cost) | March 2026 variant optimized for agent workflows — faster and cheaper per token |
| Cloud Deployments | Varies | Available through Chinese cloud providers and select international platforms |
As a fully MIT-licensed model, GLM-5 can be downloaded and deployed without any licensing fees or commercial restrictions. The primary cost consideration is compute — the full 744 billion MoE model requires substantial GPU resources, though the 8-of-256 expert activation means inference demands are closer to a 44 billion dense model.
Core Capabilities
Frontier Performance Without NVIDIA
GLM-5 claims to surpass Gemini 3 Pro on coding and agentic performance benchmarks — a remarkable achievement for a model trained on non-NVIDIA hardware:
- Coding benchmarks: Strong results across code generation, bug fixing, and multi-step programming tasks
- Agentic tasks: Designed for multi-step workflows where the model plans, executes, and iterates — not just single-turn generation
- General reasoning: Competitive with frontier models on mathematical reasoning, analysis, and knowledge-intensive tasks
MoE Architecture with DeepSeek Sparse Attention
The 744 billion parameter model uses 256 expert sub-networks with 8 activated per token:
- Efficient inference: Only ~44 billion parameters activate per forward pass, keeping compute costs closer to a mid-size model despite the massive total parameter count
- DeepSeek Sparse Attention (DSA): Borrowed from DeepSeek's architecture, DSA reduces the computational cost of attention operations — critical for handling the 200K context window efficiently
- Specialized experts: Different experts activate for different types of content, allowing deep specialization across coding, reasoning, multilingual, and domain-specific tasks
200K Context Window
The 200K token context window supports:
- Large codebase analysis: Ingest entire repositories or multi-file projects for comprehensive understanding
- Long document processing: Handle full research papers, legal contracts, and regulatory filings without chunking
- Extended agent sessions: Multi-step agentic workflows that accumulate context over many planning and execution cycles
GLM-5-Turbo for Agent Workflows
Released in March 2026, GLM-5-Turbo is optimized specifically for agentic use cases:
- Faster inference: Reduced latency for the rapid back-and-forth of agent loops
- Lower cost: Cheaper per-token pricing for high-volume agent workflows where many model calls add up
- Tool use: Enhanced function calling and structured output for integration with external tools and APIs
Strengths
- NVIDIA-free training: Proof that frontier AI can be built on domestic Chinese hardware (Huawei Ascend + MindSpore) — strategically significant for the global AI landscape
- MIT license (GLM-5, GLM 5.2 and GLM-5.3-Flash): Fully open source with no commercial restrictions, no revenue thresholds, and no competitive-use bar — among the most permissive terms at frontier scale, alongside the Apache 2.0 releases from Google (Gemma 4), OpenAI (GPT-OSS) and Meta (Muse Glimmer). The flagship GLM-5.3 is the exception in the line: its own GLM-5.3 License is MIT-shaped but adds a security-review trigger for very large model-as-a-service operators
- 744 billion MoE scale: Among the largest openly available models, with efficient 8-of-256 expert activation (~44 billion active)
- Surpasses Gemini 3 Pro: Claims to exceed Google's model on coding and agentic benchmarks
- Publicly listed company: Zhipu AI's Hong Kong IPO provides financial transparency and stability
- Tsinghua academic heritage: Research-first approach from China's top technical university
- Agent-optimized variant: GLM-5-Turbo provides faster, cheaper inference for multi-step agentic workflows
Limitations & Considerations
- Primarily optimized for Chinese: While English performance is strong, models like Llama 4 or Qwen 3.5 may offer broader multilingual coverage
- Ascend chip efficiency debate: While GLM-5 demonstrates that Huawei Ascend can train frontier models, questions remain about training efficiency compared to equivalent NVIDIA setups
- Significant compute for self-hosting: The 744 billion MoE model requires substantial multi-GPU infrastructure even though only ~44 billion parameters activate per token
- Smaller international community: Most active users and contributors are in the Chinese AI research community, with fewer English-language tutorials and integrations
- Benchmark verification: Claims of surpassing Gemini 3 Pro await broader independent validation across diverse evaluation suites
- The flagship's license is no longer plain MIT: GLM-5.3 ships under a custom GLM-5.3 License that requires a Z.ai security review before commercial use by any model-as-a-service operator whose aggregate revenue passes 10 billion dollars over twelve consecutive months. That ceiling excludes essentially every reader, but it is a real difference from the MIT terms on GLM-5, GLM 5.2 and GLM-5.3-Flash, and it means the line's licensing is now per-checkpoint rather than per-generation
- Self-hosting the flagship is a serious hardware commitment: the weights are published in FP8 at roughly 750 billion parameters, so running the top of the line locally is out of reach for anything short of a multi-accelerator server — GLM-5.3-Flash and GLM 5.2 remain the practical self-hosting options
- The cyber capability cuts both ways: the same exploitation-chain reasoning that makes GLM-5.3 strong on defensive security work is what prompted the two-week release delay. Z.ai ultimately shipped the weights rather than withholding them, which tells you what the company concluded — but a model that leads CyberGym is now downloadable by anyone, and that is the trade the open-weights argument has to carry
Best Use Cases
| Task | Why GLM-5 |
|---|---|
| NVIDIA-free AI deployment | The only frontier model trained entirely on non-NVIDIA hardware — relevant for organizations navigating chip supply constraints |
| Chinese academic research | Purpose-built for academic analysis with Tsinghua research heritage |
| Bilingual Chinese-English workflows | Native-quality output in both languages with natural code-switching |
| Open-source AI research | MIT license with full weights — maximum freedom for research and commercial use |
| Agentic workflows | GLM-5-Turbo variant optimized for fast, cost-efficient multi-step agent loops |
| Large-context analysis | 200K token window handles substantial codebases and long documents |
When to choose alternatives:
- Broader multilingual coverage beyond Chinese-English → Qwen 3.5
- Reasoning-specialist with chain-of-thought → DeepSeek R1 or DeepSeek V4
- English-first open model with largest community → Llama 4
- Commercial API with enterprise support → a closed frontier flagship from Anthropic or OpenAI
Getting Started
- Visit open.bigmodel.cn to explore the Zhipu AI platform and API documentation
- Browse the model collection on Hugging Face (THUDM) to see available model variants
- For API integration, register on open.bigmodel.cn and generate an API key — try GLM-5-Turbo first for the best balance of speed and cost
- Test a bilingual prompt (Chinese and English in the same conversation) to experience the model's code-switching fluency
- Try an agentic workflow: give GLM-5 a multi-step task (research, plan, execute) to see its planning capabilities
- For research use, review the GLM-5 technical paper to understand the Huawei Ascend training methodology and MoE architecture decisions
🎯Tip
The bigger picture: GLM-5's significance extends beyond its benchmark scores. As the first frontier model trained entirely on non-NVIDIA hardware, it represents a strategic milestone in the global AI chip competition. For organizations evaluating AI infrastructure, GLM-5 demonstrates that the NVIDIA-CUDA ecosystem, while dominant, is no longer the only path to frontier-scale AI — a development with implications for chip procurement, supply chain risk, and long-term AI infrastructure planning.
Key Takeaways
- GLM-5 is a 744 billion MoE model from Zhipu AI (Tsinghua University spinout) — the first frontier model trained entirely on Huawei Ascend chips without any NVIDIA dependency
- Released under the MIT license with 256 experts (8 activated per token, ~44 billion active), it claims to surpass Gemini 3 Pro on coding and agentic benchmarks
- Zhipu AI's January 2026 Hong Kong IPO (~$558 million raised) makes it the first publicly listed Chinese AI foundation model company
- GLM 5.2 (June 2026) extended the line to a 1 million token context window with MIT-licensed open weights — scoring 62.1 on SWE-bench Pro (ahead of GPT-5.5) and taking first place on the Design Arena leaderboard, with the GLM Coding Plan starting at $10 a month
- The current flagship is GLM-5.3 (August 14, 2026), and its weights shipped on August 28. It reuses GLM 5.2's base model and takes its gains from scaled post-training — CyberGym 77.2% to 84.5%, ExploitBench 24.4% to 54.4%, DeepSWE 46.2 to 66.9 — after Z.ai held the open weights for roughly two weeks of safety evaluation because cyber capability outgrew what it trained for
- The two checkpoints in that generation carry different licenses, and that is the lesson. GLM-5.3-Flash is plain MIT; the flagship ships under a custom GLM-5.3 License that adds a Z.ai security review for model-as-a-service operators above 10 billion dollars in revenue over twelve consecutive months — so read the LICENSE file on the repository you actually download, rather than assuming the line's permissive history covers every model in it
- Best for Chinese-English bilingual work, open-source research, and agentic workflows; its strategic importance as an NVIDIA-free proof point extends beyond any single use case
- In July 2026, Hugging Face ran GLM 5.2 locally to investigate an autonomous cyberattack after its US commercial models' guardrails blocked the forensics — an early real-world case for unrestricted open weights in defensive security
