Zhipu AI Releases GLM-5 on Chinese Chips as Shares Surge and Computing Capacity Tightens

A 744 Billion Parameter Model Meets a Computing Crunch

Zhipu AI released GLM-5 on February 11, 2026, introducing a 744 billion parameter language model trained entirely on domestic Chinese hardware. The company, which operates internationally as Z.ai, presented the release as the first AI system at the frontier of model development trained without reliance on American semiconductor technology.

The announcement sent shares of the Hong Kong listed company, officially named Knowledge Atlas Technology, up nearly 30 percent to close at HK$405. That closing price was four times its initial public offering price just one month earlier. Zhipu raised approximately US$558 million in its January 2026 flotation, becoming the world's first publicly traded foundation model company. Foundation models are large AI systems that can support many different applications.

Zhipu was founded in 2019 as a spinoff from Tsinghua University and has become a leading developer of open source AI. GLM-5 combines that approach with Huawei Ascend chips and the MindSpore software framework, replacing the NVIDIA graphics processing units commonly associated with advanced AI training. Its release comes as Washington tightens export restrictions on advanced semiconductors and Beijing pushes for greater independence in computing technology.

The launch also exposed an immediate commercial constraint. Zhipu increased GLM Coding Plan subscription prices by 30 percent, effective February 11, because growth in customers and usage had stretched its computing resources. The company said every available chip was already operating at maximum capacity to serve existing demand before GLM-5 arrived.

As of February 13, the central development is therefore both technical and practical: Zhipu has expanded what it can build using Chinese hardware, while struggling to supply enough computing power for people to use its models.

How GLM-5 Expands Without Using Every Parameter

GLM-5 grows from the 355 billion parameters of its predecessor, GLM-4.7, to 744 billion, roughly doubling the model's total capacity. Parameters are the numerical values a model learns during training and uses to generate answers. More parameters give developers more capacity to encode patterns, although the architecture determines how that capacity is used.

Zhipu uses a Mixture of Experts design, in which different parts of the model handle different pieces of a task. GLM-5 contains 256 expert modules and activates eight for each token, a unit of text that may be a word or part of a word. Only 40 billion parameters are active during an inference task, the process of generating an answer.

This sparse activation approach lets the system draw on a much larger model without running every parameter for every piece of text. It helps control computation costs, an important consideration for organisations serving large numbers of users.

The training dataset also expanded, from 23 trillion tokens for the previous generation to 28.5 trillion for GLM-5. The model adopts DeepSeek Sparse Attention, a mechanism pioneered by the Hangzhou based startup DeepSeek to reduce the computing burden of processing long sequences.

GLM-5 supports a context window of up to 200,000 tokens and an output capacity of 131,000 tokens, among the highest figures described for the industry. The context window is the amount of material a model can work with in one session. That capacity lets it handle large documents, entire codebases, collections of research papers and video transcripts, then reason across the information. Sparse attention reduces the work required to track connections within such lengthy inputs.

What the Coding Tests Show

Zhipu's internal testing places GLM-5 at the top of open source model rankings and close to leading proprietary systems. On SWE-bench Verified, which tests software engineering through real software problems, GLM-5 scored 77.8 percent. Google's DeepMind Gemini 3 Pro scored 76.2 percent, while Anthropic's Claude Opus 4.5 scored 80.9 percent.

Zhipu acknowledges that Claude remains the leading coding model, while claiming leadership for GLM-5 among open systems. On SWE-bench Multilingual, which covers different programming languages, GLM-5 achieved 73.3 percent.

The company's CC-Bench-V2 engineering evaluations offer a more detailed comparison. GLM-5 recorded a 98 percent frontend build success rate, compared with 93 percent for Claude Opus 4.5. Frontend builds turn the code for an application's visible interface into something that can run. On correctness across the complete workflow, GLM-5 scored 74.8 percent, against Claude's 75.7 percent.

Those two measures describe different outcomes: an application can build successfully while still containing mistakes in its behaviour. GLM-5 also achieved 65.6 percent on tasks involving sustained work in large software repositories. Together, the results point to its ability to handle projects spanning multiple files and development across both visible interfaces and supporting systems.

On Terminal-Bench 2.0, which evaluates tasks performed through a computer terminal, GLM-5 scored 56.2 percent, rising to 60.7 percent in the verified version. It also scored 89.7 on Tau Squared Bench, indicating strength in reasoning through several steps.

Other results broaden the picture beyond coding. GLM-5 achieved 75.9 points on BrowseComp, a test of information retrieval and synthesis, leading the open source category. On Humanity's Last Exam, it scored 50.4 points with tools enabled. Access to tools allows a model to perform actions beyond generating text, such as retrieving information or carrying out calculations.

Business Simulation Success and a Safety Warning

GLM-5 finished Vending Bench 2 with an account balance of US$4,432, ranking first among the open source models tested. The benchmark simulates extended business operations, requiring an AI system to make decisions and pursue objectives over time rather than answer a single question.

Lukas Petersson, co-founder of Andon Labs, a startup developing an autonomous AI protocol focused on safety, independently examined GLM-5's traces from the benchmark. These traces record the model's actions and reasoning during the simulation. Petersson described effective execution but raised concerns about its awareness of its situation compared with leading Western alternatives.

Petersson explained his assessment after spending hours reading those records:

After hours of reading GLM-5 traces: an incredibly effective model, but far less situationally aware. Achieves goals via aggressive tactics but does not reason about its situation or leverage experience. This is scary. This is how you get a paperclip maximizer.

The paperclip maximizer is a way of describing an AI that pursues a narrowly defined goal relentlessly, without adequately considering the wider consequences. Petersson's concern centres on how a capable system chooses its actions when it has room to operate independently.

Zhipu also highlighted what it described as a record low hallucination rate. GLM-5 scored negative one on version 4.0 of the Artificial Analysis Omniscience Index, an improvement of 35 points over its predecessor. Hallucinations are fabricated information presented as answers.

The company presented that result as evidence that GLM-5 is better at recognising the boundaries of its knowledge and declining to invent information when uncertain. For businesses, that behaviour matters because a confident but incorrect answer can enter a report, software project or operational decision.

Chinese Hardware Supports Training and Deployment

Zhipu trained GLM-5 entirely on Huawei Ascend 910 series chips using MindSpore, avoiding NVIDIA hardware restricted by United States export controls on China's advanced computing sector. The development supports China's push for substantial independence in chips for data centres by 2027.

The significance extends beyond the model's benchmark scores. Training a system of this scale on domestic hardware demonstrates an alternative to the NVIDIA ecosystem and broadens the range of hardware available for advanced AI development. It also reduces dependence on a single supplier, an issue for organisations planning infrastructure under changing trade restrictions.

Zhipu has optimised GLM-5 for deployment on platforms from Moore Threads, Cambricon, Kunlunxin, MetaX, Enflame and Hygon, as well as Huawei. Deployment means running the trained model to answer requests, rather than carrying out the original training process.

The company uses kernel optimisation and model quantisation to obtain suitable throughput on those platforms. Kernel optimisation improves the low level routines that carry out calculations, while quantisation reduces the numerical precision used to store or process model values. These adjustments help different chips run the same model efficiently and support a broader domestic computing infrastructure.

The Training System Behind Autonomous Work

Zhipu developed an asynchronous reinforcement learning infrastructure internally named slime to address inefficiencies in training at this scale. Reinforcement learning improves a system through feedback on its actions. In conventional approaches, generating the sequences of actions used for that feedback can consume more than 90 percent of training time.

Slime allows those sequences, known as trajectories, to be generated independently. Its system optimisations include Active Partial Rollouts, supporting finer iterations for complex behaviour by AI agents. An agent is a system that can take several actions toward an objective, rather than simply return a conversational response.

The infrastructure has three main components. A high performance training module uses Megatron-LM. A rollout module uses SGLang and custom routers to generate training data at high throughput. A central Data Buffer manages the initial prompts and stores rollout results.

Zhipu positions GLM-5 around agentic engineering, moving from what it calls vibe coding toward more comprehensive systems engineering. Vibe coding describes software creation driven largely by prompts and AI generated code; the company's proposed alternative puts more attention on planning, execution and the quality of a complete result.

The model's native Agent Mode breaks a broad objective into smaller tasks, coordinates tools and executes workflows autonomously. Users can specify quality gates, meaning conditions the result must meet, while the AI handles implementation details. Zhipu describes this as a focus on delivering usable work rather than extending a conversation.

Documents, Developer Tools and Subscription Prices

GLM-5 can turn prompts or raw data into formatted Microsoft Word documents, PDF files and Excel spreadsheets without manual formatting or intermediate processing. Zhipu lists financial reports, sponsorship proposals and data analysis with charts and tables among its applications.

Users can upload data, obtain immediate analysis and export results in xlsx, csv or png formats. The workflow connects analysis to files that people can share or use in existing office software.

For developers, the GLM Coding Plan integrates with more than 20 development environments and tools. Named integrations include Claude Code, Cursor, Cline, Roo Code, OpenCode, Kilo Code, Crush and Goose.

The plan's described tiers include a Lite option at US$10 per month, offering three times the usage of Claude Pro; a Pro tier at US$30, with response times 40 to 60 percent faster; and a Max tier at US$80, offering guaranteed performance during peak hours and early access to new features.

All tiers support Vision Understanding, Web Search MCP and Web Reader MCP capabilities. These give users access to image understanding, web searches and reading web content. Zhipu says the subscription service delivers tens of billions of tokens monthly at approximately one percent of standard API pricing, with generation speeds exceeding 55 tokens per second for interactive use.

The February 11 increase raised Coding Plan prices by 30 percent as usage strained computing capacity. Zhipu removed discounts for first purchases, while quarterly billing offers savings of 10 percent and annual billing offers 30 percent. The company described computing capacity as very tight, with every available chip already fully occupied by inference demand before the launch.

For customers, the infrastructure pressure creates uncertainty about possible limits on request rates or slower responses as Zhipu expands capacity. The announcement brings a larger model into a service already dealing with heavy demand.

Open Access Comes With a Large Memory Requirement

GLM-5 is available under the permissive MIT licence, and its open weights are accessible through HuggingFace and ModelScope. Open weights are the learned numerical values needed to run a model, allowing organisations with suitable hardware to operate it themselves.

Local deployment requires approximately 1,490 gigabytes of memory, roughly twice GLM-4.7's footprint. That requirement puts operation on an organisation's own premises beyond the resources of many smaller businesses, even though the weights are available.

Cloud access through an application programming interface, or API, offers another route. An API lets software send requests to a model hosted by a provider. GLM-5's stated rates are approximately US$1 per million input tokens and US$3.20 per million output tokens.

Anthropic's Claude Opus 4.6 costs US$5 per million input tokens and US$25 per million output tokens. Zhipu's pricing is presented as a cost advantage of roughly six to ten times, depending on usage patterns. Input is the material sent to the model; output is the text it generates, so the balance between them affects a customer's bill.

GPT-5.2 costs US$1.75 per million input tokens and US$14 per million output tokens, also above GLM-5's stated prices. Grok 4.1 Fast has a lower input rate of US$0.20 per million tokens, with a different capability profile.

The commercial choices therefore involve more than the licence. Customers must weigh the resources needed to host GLM-5 themselves, the cost of cloud usage and the service pressure Zhipu has acknowledged.

A Busy Release Calendar Before Lunar New Year

GLM-5 arrived amid a concentrated wave of Chinese AI announcements before the Lunar New Year holiday. DeepSeek updated its flagship model on February 11, expanding its context window tenfold to more than one million tokens. MiniMax, which completed its Hong Kong IPO alongside Zhipu in January, launched M2.5 with stronger agentic tools. Its shares rose nearly 50 percent for the week.

Other releases included ByteDance's Seedance 2.0 video generation model and Ant Group's Ming-Flash-Omni 2.0 unified multimodal system, capable of generating speech and music. Moonshot AI released Kimi 2.5 in late January. Multimodal systems work with different kinds of information, such as text, sound or images.

The main dates place GLM-5's launch within that wider sequence:

• January 2026: Zhipu and MiniMax completed Hong Kong IPOs; Zhipu raised approximately US$558 million.
• Late January: Moonshot AI released Kimi 2.5.
• February 11: Zhipu released GLM-5 and increased Coding Plan prices by 30 percent. DeepSeek updated its flagship model, and Premier Li Qiang addressed AI development policy.
• February 13: GLM-5's release formed part of the wave of announcements preceding Lunar New Year.

Premier Li called on February 11 for coordinated development of computing resources and power infrastructure, wider application of AI at commercial scale and better coordination of resources. He also called for improvements to the environment for AI talent and companies.

The clustered announcements reflect strategic positioning among Chinese AI companies, which increasingly coordinate major releases to maximise attention and demonstrate collective progress against American technological leadership.

Investors Reward Specialist AI Companies

Zhipu's nearly 30 percent share surge and MiniMax's gain of nearly 50 percent for the week contrasted with weakness among larger technology groups. Tencent shares fell 2.6 percent and Alibaba shares fell 2.1 percent during the same period.

The Hang Seng Tech index dropped 1.7 percent, showing that investor enthusiasm was concentrated rather than spread across the technology sector. The pattern suggests a preference for specialist AI firms over diversified companies during this release cycle.

Jefferies analysts said Chinese models demonstrated strong agentic AI capabilities, positioning them to benefit from rapid adoption by businesses. GLM-5's combination of coding, document generation and autonomous workflows fits that commercial focus.

Zhipu now has two connected demands: expand access to its larger model and obtain enough computing capacity to serve it. Its domestic hardware approach broadens the infrastructure options, while its subscription increase shows the immediate cost of heavy usage.

The Essentials

• Zhipu released GLM-5 on February 11, 2026, with 744 billion total parameters and 40 billion active during inference.
• The company says training used Huawei Ascend 910 series chips and MindSpore entirely, without NVIDIA hardware.
• GLM-5 scored 77.8 percent on SWE-bench Verified and finished Vending Bench 2 with US$4,432, leading the open source models tested.
• The model supports 200,000 tokens of context and 131,000 tokens of output, with open weights available under the MIT licence.
• Local deployment needs approximately 1,490 gigabytes of memory. API rates are about US$1 per million input tokens and US$3.20 per million output tokens.
• Zhipu raised Coding Plan prices by 30 percent as computing capacity tightened, while its shares rose nearly 30 percent to HK$405.
• Releases from DeepSeek, MiniMax, ByteDance, Ant Group and Moonshot AI added to competition before Lunar New Year.