Meet the ‘Chinese Claude Code’ Looking to Create the Next DeepSeek Moment

Even though American companies are leading the AI race with new coding models, China is not too far behind, already flexing its AI muscles with open-source models like DeepSeek. It’s betting on six “AI Tigers” to drive the country’s generative AI race, backing them with state capital for rapid innovation.

One of those AI tigers is Z.ai, formerly known as Zhipu AI. The Beijing-based company has launched GLM-5, a 744-billion-parameter foundation model built for complex systems engineering and long-horizon agentic tasks. The model positions itself as a formidable challenger to leading Western systems such as OpenAI’s GPT-5.2 and Anthropic’s Claude Opus 4.5.

The company said GLM-5 increases total parameters from 355 billion used in GLM-4.5 and 4.5-Air to 744 billion, with active parameters rising from 32 billion to 40 billion. Pre-training data has expanded from 23 trillion to 28.5 trillion tokens. The model also integrates DeepSeek Sparse Attention (DSA), which the company said reduces deployment cost while maintaining long-context capability.

On coding benchmarks, GLM-5 scored 77.8 on SWE-bench Verified, outperforming Gemini 3 Pro at 76.2, while trailing Claude Opus 4.5, which scored 80.9. On Terminal Bench 2.0, it posted 56.2, narrowing the gap with Claude Opus 4.5 at 59.3. In general agent performance, GLM-5 scored 75.9 on BrowseComp with context management, surpassing Claude Opus 4.5 on that benchmark.

In an exclusive interview with AIM, Lou, Head of Developer Ecosystem and Partnerships at Z.ai, who goes by her first name, says GLM-5 marks a milestone for China’s open-source AI ecosystem. 

“GLM-5 is shaping up to be the closest open-source model to Opus 4.5,” she says. “A lot of people outside of China don’t know us. That’s why I think we might create the next DeepSeek moment.” 

China’s AI industry drew global attention following DeepSeek’s breakthroughs in open-source reasoning models. Z.ai believes GLM-5 could replicate that impact, this time in coding.

Lou acknowledges that Opus 4.5 has set a high bar in reasoning and coding globally. However, the AI company is not claiming outright superiority yet. “We don’t really say we want to compete at that level yet. We’re still catching up, but we think we’re on the same path,” she says.

Pricing plays a key role in Z.ai’s strategy, especially as access to leading Western models remains restricted or costly for developers in China. “A lot of people in China cannot access Claude Code, and it’s really expensive,” Lou says, adding that developers in China consider Z.ai to be “the Chinese Claude Code.”

Features and Availability

Z.ai claims GLM-5 delivers “best-in-class performance among all open-source models in the world on reasoning, coding, and agentic tasks, closing the gap with frontier models.” The model has been open-sourced on open-source communities Hugging Face and ModelScope under the MIT License. 

Developers can access GLM-5 through Z.ai’s GLM Coding Plan and integrate it with popular coding agents, including Claude Code, OpenCode, Kilo Code, Roo Code, Cline, and Droid. The rollout is being phased due to limited compute capacity and advanced chip restrictions due to US export controls. 

Initial access is available to Max-tier subscribers, costing $216 per quarter. The Pro plan is expected to complete integration by February 16. GLM-5 is available on Z.ai’s platform in both Chat and Agent modes.

The model can generate .docx, .pdf and .xlsx files directly from prompts, including product requirement documents, lesson plans, financial reports, and spreadsheets. 

Z.ai said it is rolling out Agent mode within its app, with built-in tools for Word, PDF and Excel creation, supporting multi-turn collaboration.

The company said GLM-5 supports local deployment via vLLM and SGLang and can run on non-NVIDIA chips—including domestic alternatives like Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame, and Hygon—through kernel optimisation and quantisation.

If DeepSeek marked China’s reasoning breakthrough, Z.ai hopes GLM-5 will define its coding chapter. “We want to lead a coding momentum. I think 2026 is the year of coding,” Lou says

IPO Momentum and Global Expansion

Founded in 2019 by Tsinghua University researchers, including Tang Jie and Li Juanzi, Z.ai listed on the Hong Kong Stock Exchange on January 8 this year under ticker 2513.HK, becoming the world’s first publicly-listed foundation model company. It raised ~$560 million at a $6.6 billion valuation.

Upon the release of GLM-5, its shares surged over 28% on the Hong Kong Stock Exchange. 

Before its public debut, the company, then Zhipu AI, secured $140 million in funding from Shanghai’s state-backed investors.

The company is reported to have offices in the United Kingdom, Singapore, Malaysia, and several countries in the Middle East. It also runs joint innovation center projects across Southeast Asia, including in Indonesia and Vietnam.

Chinese AI Push 

Z.ai’s push does not come in isolation. Chinese AI startups are launching new models in quick succession. 

On January 27, Alibaba-backed Moonshot AI unveiled Kimi K2.5, its latest open-source model featuring vision-based coding and a self-directed agent swarm system. The model combines text and visual inputs within a unified multimodal architecture, allowing it to handle reasoning, coding, visual understanding, and autonomous task execution in a single system.

Lou says that in China, Z.ai is already being compared with domestic competitors, including Kimi. She argues that while Kimi excels in interface design, it is not built specifically for software engineers.

“Kimi’s feature is aesthetic. They have an amazing UI, but it’s not really for software engineers,” she claims. Z.ai, by contrast, sees itself as the coding-first alternative.

Meanwhile, ByteDance’s latest video generation model, Seedance 2.0, has taken the internet by storm, with users creating anime-style videos using the tool.

The model stands out for its multi-shot storytelling, smooth scene transitions, character consistency across frames, and realistic physics such as natural motion and collision effects. It also delivers 2K resolution outputs, and the company claims it’s 30% faster than its predecessor.

MiniMax AI has expanded access to M2.5, making the model instantly available across its web and desktop Agent platforms.

Post GLM-5, all eyes are now on DeepSeek’s upcoming V4 model, which industry watchers expect to be the company’s next major release.