Zhipu AI released GLM-5 on Wednesday, a 745-billion-parameter model that the company says approaches Anthropic's Claude Opus 4.5 on coding benchmarks and surpasses Google's Gemini 3 Pro on others. The Beijing-based company, which listed on the Hong Kong Stock Exchange last month in a $558 million IPO, is pitching GLM-5 as its most capable model to date, with a particular focus on coding and long-running agent tasks.
The "approaches Claude Opus" framing is doing some heavy lifting here. Zhipu's own press release makes the comparison, and without independent benchmarks published yet, that claim is hard to evaluate. The company's previous model, GLM-4.7, was competitive with Claude Sonnet 4.5 on several coding benchmarks but not dominant, so "approaching Opus" would represent a meaningful jump if true.
The architecture, and what's borrowed
GLM-5 uses a Mixture of Experts architecture with roughly 745 billion total parameters and 44 billion active per inference. That's about double the scale of GLM-4.5, which ran 355 billion total with 32 billion active. The company also adopted DeepSeek Sparse Attention, a mechanism DeepSeek developed for its V3.2 model that selectively reduces which tokens the model attends to during long sequences. It's an efficiency trick, not a performance trick, and it lets GLM-5 handle longer contexts without the compute costs spiraling.
Using a competitor's published research is perfectly normal in open-source AI. But it's worth noting: Zhipu is borrowing architectural components from DeepSeek while simultaneously competing for the same "top Chinese AI lab" positioning. DeepSeek, for its part, released DSA as part of its V3.2 technical report, seemingly unconcerned about who picks it up.
The chip story matters more than the model
The more significant detail might be buried in Zhipu's press release. GLM-5 was developed using domestically manufactured chips for inference, including Huawei's flagship Ascend processors alongside hardware from Moore Threads, Cambricon, and Kunlunxin. According to Reuters reporting, the model was trained using MindSpore, Huawei's machine learning framework.
Zhipu already proved this was possible on a smaller scale. In January, the company released GLM-Image, an image generation model trained entirely on Huawei's Ascend Atlas 800T A2 servers. Scaling that approach to a 745B language model is a different challenge, and Zhipu hasn't disclosed how many processors or how long training took. Those details matter: training on domestic Chinese chips is a solved problem; training efficiently is not.
Beijing wants these announcements. The Chinese government has been pushing AI companies to demonstrate that U.S. chip export restrictions haven't crippled their progress, and a model release timed to the Lunar New Year holiday provides exactly the kind of visibility officials want.
What Zhipu isn't saying
There are no detailed benchmarks yet. The company claims GLM-5 will be open-source, which would follow its pattern with GLM-4.7 (currently available on Hugging Face under an MIT license). But "will be" and "is" are different things, and the model apparently won't fully roll out until Thursday.
Zhipu also hasn't addressed pricing, context window specifics beyond the rumored 200,000 tokens, or how GLM-5 performs outside of coding tasks. The company has consistently positioned its models around coding and agentic capabilities, which is smart branding but makes it hard to assess general-purpose performance.
The timing is part of a broader Chinese AI release wave. ByteDance dropped Seedance 2.0 last week, Kuaishou launched Kling 3.0 days before that, and DeepSeek reportedly updated its own model on the same day as GLM-5. For a company that went public at a $7.1 billion valuation just five weeks ago while reporting net losses of 2.36 billion yuan in the first half of 2025, the pressure to keep shipping is obvious.
GLM-5 is expected to go live on Zhipu's chat.z.ai platform on Thursday. Model weights should follow on Hugging Face and ModelScope, though Zhipu hasn't committed to a timeline.
