by Tan Aik Keong (AK)
As one of China's "AI Tiger" startups, Zhipu AI (also known as Zhipu Z.ai) has racked up a string of frontier results in large language models in recent years. Spun out of Tsinghua University and founded by Zhang Peng and Tang Jie's team in 2019, Zhipu has focused from the start on general-purpose large language models — and in just a few years, has built several genuinely milestone models.
A track record of firsts
GLM-130B — the world's first open-source, bilingual (Chinese-English) 130-billion-parameter pretrained model, released in 2022. In Stanford University's 2022 evaluation of 30 leading global large models, GLM-130B was the only Asian model selected, and performed strongly — matching OpenAI's GPT-3 175B base model on key accuracy metrics, real proof a Chinese team could keep pace with the world's best on frontier NLP work.
The ChatGLM series — conversational models built on GLM-130B. In 2023, Zhipu released the hundred-billion-parameter ChatGLM and open-sourced a lightweight version, ChatGLM-6B. Despite having just 6.2 billion parameters, careful fine-tuning let it match the conversational quality of much larger models, making it one of the first large models an ordinary developer could realistically deploy locally. ChatGLM passed a million downloads globally within a month of going open source, lowering the barrier for the public to try a homegrown conversational AI.
GLM-4 and GLM-4.5 — Zhipu's next-generation general models, released in 2024 and 2025 respectively. GLM-4 launched in early 2024 and ranked second for the year on the Shanghai AI Laboratory's authoritative OpenCompass leaderboard, behind only GPT-4 Turbo — the top-ranked open-source model domestically, with overall performance closing in on GPT-4. In July 2025, Zhipu followed up with GLM-4.5, a 355-billion-parameter sparse model, released fully open source. It's the first single model to natively combine complex reasoning, code generation and tool-using agentic capability — in short, GLM-4.5 doesn't just chat, it can reason independently, write code, and act as an "AI employee" completing multi-step tasks. That kind of autonomous agentic capability marks a real step toward general intelligence.
Briefly the most-talked-about model in the world
Zhipu's strength shows up in the leaderboards too. GLM-4.5 topped the global trending chart on Hugging Face, the world's largest open-source AI community, just two days after release — climbing to second place within 10 hours of its 28 July launch and first place within two days, setting a growth record on the platform. That's a real signal of how much attention and credibility Zhipu's open-source models command globally.
The performance holds up too. By Zhipu's own account, GLM-4.5 ranked third globally across 12 authoritative benchmarks spanning graduate-level reasoning and complex software engineering — first among all Chinese and open-source models. In other words, outside a handful of top proprietary models like GPT-4, GLM-4.5's average performance sits in the same tier as the world's strongest AI systems. In early August, Zhipu also released a multimodal version, GLM-4.5V, which broke more than 40 global records in visual reasoning and has been called, by some in the industry, an "almost unbeatable" open-source multimodal model. Taken together, Zhipu's large-model technology now sits genuinely in the global first tier.
Zhipu versus DeepSeek
Among China's AI upstarts, another name draws just as much attention: DeepSeek. Founded in 2023, this startup made its name with low-cost, high-performance large models. Earlier this year, DeepSeek released its open-source DeepSeek-R1, training a model that rivalled GPT-4 with less than half the compute of major international labs — a shock to the industry. Some analysts described R1 as proof that "low cost can still build top-tier AI," and in some quarters it was framed as a "Chinese Sputnik moment," briefly worrying the US that its lead in the AI race was closing faster than expected.
DeepSeek's core edge is efficiency — through innovative training optimisation, it lifted compute utilisation even without access to top-tier chips, achieving results close to the strongest models at a fraction of the scale and cost. DeepSeek-R1 also released its model weights openly, letting developers worldwide freely use and improve on it, further amplifying its reach.
Facing that competition, Zhipu has chosen to compete head-on with its own distinct approach. GLM-4.5 also stays open source and pushes parameter efficiency hard — at half DeepSeek-R1's parameter count, it still outperforms it on several standard benchmarks. In other words, Zhipu achieved a performance edge with a "smaller" model, setting a new frontier on the performance-to-scale ratio. GLM-4.5 also uses a Mixture-of-Experts (MoE) architecture to substantially cut inference overhead, keeping running costs well below mainstream closed-source models — its API pricing runs as low as ¥0.8 per million input tokens and ¥2 per million output tokens, a fraction of what some Western commercial models charge. On both performance and cost, Zhipu is closing the gap with leaders like DeepSeek fast, and in some areas has already pulled ahead.
More importantly, the two companies differ in technical direction. DeepSeek's focus is making models "light" and "cheap," known for high efficiency on limited resources; Zhipu is pursuing something more comprehensive, integrating multiple capabilities into a single model. GLM-4.5, for instance, doesn't just answer questions — it can independently call a browser, a code interpreter and other tools to complete complex tasks. CNBC has noted that new models from Chinese AI companies keep getting both smarter and cheaper to run — echoing exactly the breakthrough that made DeepSeek's debut so notable. DeepSeek pioneered the cost-and-efficiency breakthrough; Zhipu, building on that same efficiency, has pushed further to broaden what a large model can actually do.
On OpenAI's radar
Zhipu's rise hasn't just caught users' and the media's attention — it's caught the attention of the world's AI giants too. OpenAI, the company behind ChatGPT, has publicly name-checked Zhipu directly. In late June 2025, OpenAI published an analysis of China's AI industry noting that its analysts had observed Zhipu winning significant government contracts across multiple countries — a sign, in OpenAI's own assessment, of China's growing momentum in the global AI race. OpenAI noted Zhipu is actively working to export Chinese AI systems and standards to emerging markets, positioning itself as a "responsible, transparent, auditable" Chinese AI alternative. That mention made clear Zhipu has already moved well beyond China's borders, actively expanding across Southeast Asia, the Middle East and Africa. Reuters has reported that Zhipu has drawn even more attention since OpenAI's mention.
Some seasoned industry watchers have noted that while DeepSeek gets more media attention, the quieter, more pragmatic Zhipu may actually be the bigger competitive threat in OpenAI's own eyes. OpenAI CEO Sam Altman has repeatedly stressed that Chinese AI capability can't be ignored, and Zhipu — one of the few Chinese unicorns with both frontier model development and genuine international expansion capability — is widely seen as one of the strongest contenders for the "Chinese OpenAI" title. The US government even added Zhipu AI to its export-control Entity List this past January, restricting its access to advanced US components — in some readings, a sign of how seriously the US takes Zhipu's technical strength and global ambitions.
Undeterred by the pressure
Encouragingly, international competitive pressure hasn't slowed Zhipu down. If anything, the company has leaned further into open collaboration to meet the challenge — continuing to open-source its newest, strongest models for the global developer community to build on, while actively pursuing overseas partnerships and capital. Reports suggest Zhipu is considering a Hong Kong listing as early as 2025 to help fund its global expansion. Taken together, Zhipu AI is no longer a quiet also-ran — it's a fast-rising leader on the world AI stage.
Closing: a contender building real momentum
Overall, Zhipu AI is rewriting perceptions of Chinese AI companies through solid technical innovation and a genuinely open, collaborative approach. From performance that rivals the world's best models, to topping global leaderboards, to drawing direct attention from a giant like OpenAI, Zhipu has backed up its ambitions with results. The name may not yet be as familiar to the general public as Baidu or Alibaba, but within AI circles, it's already regarded as one of China's — and the world's — most promising AI dark horses. As Zhipu CEO Zhang Peng has put it: building a powerful, genuinely independent and controllable foundation model isn't something that can wait. Looking ahead, with sustained R&D investment and genuine innovative drive, Zhipu AI looks well placed to claim real ground on the path toward general AI — doing credit to Chinese technology, and injecting fresh energy into the global AI industry. Once underrated, this player is fast becoming one that shapes the rules of the game.
Part of the AK AI Corner column. Originally published in Oriental Daily (东方日报) on Sep 6, 2025.