Remember when everyone assumed you needed NVIDIA's best GPUs to build a world-class AI model? Zhipu AI just challenged that assumption in a big way.
The Beijing-based startup (trading on the Hong Kong Stock Exchange as Knowledge Atlas, ticker 2513) released GLM-5 this week, a 744-billion-parameter open-source model that approaches Anthropic's Claude Opus on coding benchmarks and surpasses Google's Gemini 3 Pro on several others. The kicker: it was trained entirely on Huawei Ascend chips using the MindSpore framework. Zero American hardware involved.
That's a sentence worth re-reading.
First up, the TL;DR
If you only have 2 mins, read this:
Zhipu AI has released GLM-5, a massive 744-billion-parameter (a.k.a size of the AI brain) open-weights model (meaning anyone with enough GPUs can run the model themselves; usually you do this over the cloud) that challenges the assumption that top-tier AI requires US hardware.
Key Takeaways
- Hardware Independence: The model was trained entirely on Chinese chips (Huawei Ascend) using the MindSpore framework, proving that US export controls on NVIDIA GPUs did not prevent China from building a frontier model.
- Performance: It is highly competitive in coding, scoring 77.8% on SWE-bench (approaching Claude Opus 4.5's 80.9%) and performing well in graduate-level science reasoning.
- Architecture: It uses a Mixture of Experts (MoE) design. While total parameters are 744B, only ~40B are active per task, keeping it efficient.
- The Trade-off: While powerful and significantly cheaper to run than GPT-5 or Claude (roughly 1% of standard API pricing for coding plans), it is slower, generating roughly 17–19 tokens per second compared to the 25–30+ of its competitors.
Why It Matters This signals a shift in the "AI Arms Race." It demonstrates that Chinese companies can produce world-class, open-source models using domestic hardware, offering a viable, lower-cost alternative to Western developers and emerging markets.
Resources:
What GLM-5 Actually Is
GLM-5 uses a Mixture of Experts (MoE) architecture, the same approach behind DeepSeek's models. Think of it like a company with 256 specialists on staff, but only 8 clock in for any given task. The result: 744 billion total parameters, but only about 40 billion are "active" at any moment, which keeps the model fast and (relatively) affordable to run.
Here are the headline specs:
- 744B total parameters (up from 355B in its predecessor GLM-4.5), with 40B active per inference
- 200K token context window with a 131K max output (one of the longest in the industry)
- 28.5 trillion training tokens (up from 23T in GLM-4.5)
- Open-source under MIT license on HuggingFace, meaning anyone can download, fine-tune, and commercially deploy it
- Built using DeepSeek Sparse Attention (DSA), a technique that keeps long-context processing efficient without the usual computational overhead
How It Stacks Up
According to Artificial Analysis, GLM-5's Reasoning variant scores a 50 on their Intelligence Index (the average is 25), placing it among the leading models globally. But the nuance matters.
Where GLM-5 shines:
- Coding: 77.8% on SWE-bench Verified, approaching Claude Opus 4.5's 80.9%. It also leads all open-source models on Vending Bench 2 and BrowseComp.
- Math and science reasoning: Competitive scores on GPQA Diamond (graduate-level science) and AIME 2025 (math competition benchmark).
- Agentic tasks: Built-in tool use, web browsing, and multi-step task execution. Zhipu says this marks a shift from "vibe coding" to what they call "agentic engineering."
Where it falls short:
- Speed: 17-19 tokens per second on throughput, compared to 25-30+ for competitors. It's noticeably slower.
- English performance: Most benchmarks so far skew toward Chinese-language tasks. Independent English evaluations are still rolling in.
- Claude still leads on coding overall. South China Morning Post confirmed that GLM-5 "still lagged Anthropic's Claude" across the board on coding benchmarks, per Zhipu's own self-reported scores.
The Price Gap Is Wild
This is where it gets interesting for developers. GLM-4.x API pricing sits at roughly $0.11 per million tokens. For context, GPT-5 charges $1.25 per million input tokens and $10 per million output tokens. Claude Opus is even pricier.
GLM-5 pricing hasn't been finalized, but if it follows the trend of its predecessors, you're looking at frontier-level capability at a fraction of what American models charge. For cost-sensitive applications (think startups, developers in emerging markets, or high-volume enterprise use cases), that's a meaningful difference.
Zhipu also launched a dedicated GLM Coding Plan subscription that gives developers access across mainstream coding tools at roughly 1% of standard API pricing. We haven't tested it yet, but the pricing alone is worth watching.
The Chip Story Is the Real Story
US export controls on advanced semiconductors were supposed to slow China's AI progress. The logic: if you can't buy NVIDIA's best GPUs, you can't train frontier models.
GLM-5 is a direct counter-argument. Zhipu trained the entire model on Huawei Ascend chips alongside chips from Moore Threads, Cambricon, and Kunlunxin, all Chinese manufacturers. Beijing has been pushing domestic companies to prove this is possible, and Zhipu just provided the most high-profile evidence yet.
Does this mean Chinese chips have caught up with NVIDIA? Not exactly. The throughput numbers suggest training was likely less efficient and took longer. But "less efficient" and "impossible" are very different things.
As Bloomberg reported, the release is designed to "jolt" the competitive landscape, and it arrives alongside a wave of Chinese model launches ahead of Lunar New Year: ByteDance's Seedance 2.0, Kuaishou's Kling 3.0, and others are all dropping within days of each other.
The "Pony Alpha" Easter Egg
Fun backstory: a mysterious model called "Pony Alpha" appeared on OpenRouter in early February, posting coding benchmarks that rivaled Claude Opus. The AI community quickly traced it back to Zhipu through GitHub pull requests and benchmark analysis. Zhipu has since confirmed the connection. Stealth launches are apparently a global phenomenon now.
Why This Matters
For most Neuron readers, GLM-5 probably won't replace Claude or ChatGPT in your daily workflow tomorrow. The English-language performance data is still incomplete, and the speed tradeoffs are real.
But the bigger picture is significant. The era where only a handful of American companies could build frontier AI models is over. GLM-5 is open-source, competitive on benchmarks, and built on hardware that the US tried to restrict. That changes the calculus for everyone: developers who want cheaper alternatives, companies evaluating vendor lock-in, and policymakers who assumed export controls would maintain a capability gap.
Zhipu raised $558M in its Hong Kong IPO last month and its stock has rallied. Expect the company to push hard on international expansion, especially in Southeast Asia and Belt & Road nations where cost-competitive open-source models have an obvious appeal.
If you want to try GLM-5 yourself, it's available on Z.AI's platform, through OpenRouter, and the full weights are on HuggingFace under MIT license.
