Zhipu AI Launches GLM‑4.5, an Open-Supply 355B AI Mannequin Aimed toward AI Brokers

Chinese language startup Zhipu AI (now rebranded as Z.ai) has open-sourced its new flagship mannequin GLM‑4.5, a 355-billion-parameter basis AI mannequin. The July 28 announcement touts GLM-4.5’s state-of-the-art efficiency in reasoning, coding, and autonomous “agent” duties, positioning it as a homegrown challenger to the likes of OpenAI’s GPT-4.

A New Flagship Mannequin Born for AI Brokers

Zhipu’s GLM-4.5 is explicitly designed for agent functions – AI methods that may perform autonomous duties by interacting with instruments and environments. It’s billed as the primary “SOTA-level” (state-of-the-art) native agent mannequin in China . Not like commonplace chatbots, GLM-4.5 natively integrates complicated logical reasoning, code era, and interactive decision-making capabilities inside a single mannequin . This marks the primary time Zhipu has fused a number of core talents into one AI system, reflecting a technical breakthrough geared toward autonomous AI brokers .

Technically, GLM-4.5 adopts a Combination-of-Specialists (MoE) structure, enabling an unlimited mannequin scale with improved effectivity. The total GLM-4.5 has 355 billion parameters in whole, of which about 32 billion are lively for any given question . Zhipu additionally gives a lighter GLM-4.5-Air variant (106 billion whole, 12 billion lively parameters) for prime effectivity deployments . This MoE design permits the mannequin to realize excessive efficiency with out the runtime price of using all parameters without delay, successfully doubling parameter effectivity in comparison with earlier fashions .

The mannequin was educated on an unprecedented dataset – 15 trillion tokens of common textual content pre-training, adopted by 8 trillion tokens of focused fine-tuning information for code, reasoning, and agent conduct, with extra reinforcement studying to boost these abilities . GLM-4.5 additionally helps an prolonged 128,000-token context window, enabling it to deal with very lengthy inputs or multi-step interactions, far exceeding typical context lengths of fashions like GPT-4 (8k–32k) and even Claude 2 (100k)  . Zhipu launched a dual-mode operation: a “considering mode” for complicated reasoning and power use, and a “quick response” mode for rapid solutions . Builders can toggle this to steadiness depth of reasoning towards velocity, which is especially helpful for agent use-cases that generally require chain-of-thought planning .

Benchmark Efficiency: Among the many World’s Finest

Preliminary evaluations place GLM-4.5 among the many top-tier of AI fashions globally. On a collection of 12 consultant benchmarks masking information check (MMLU Professional), arithmetic (MATH500), coding (LiveCodeBench), multi-step reasoning (TAU-Bench), and extra, GLM-4.5 achieved an general rating rating within the prime three worldwide  . This makes it the highest-ranking mannequin amongst Chinese language-developed and open-source fashions, narrowing the hole with solely the perfect proprietary fashions. Zhipu even experiences that by the typical of those assessments, GLM-4.5 is #2 globally, successfully on par with the world’s strongest AI methods like GPT-4 . Chinese language media famous the mannequin’s efficiency – similar to the highest flagship AI fashions on the planet .

Notably, GLM-4.5 shines in coding and “agentic” duties. In Zhipu’s inner assessments simulating software program improvement situations, GLM-4.5 was benchmarked towards main coding-focused fashions comparable to Anthropic’s Claude Code, Moonshot’s Kimi-K2, and Alibaba’s Qwen-3 Coder. GLM-4.5 demonstrated superior job completion and extra dependable instrument utilization than different open fashions, and got here very near Anthropic’s newest Claude-4 (Sonnet version) on many programming challenges . Zhipu acknowledges Claude-4 nonetheless holds a slight edge on sure dimensions, however says GLM-4.5 can already function a viable drop-in various for many actual improvement situations  . To again this, the corporate publicly launched the 52 coding duties and agent interplay logs used within the analysis for group scrutiny .

Parameter-for-parameter, GLM-4.5 is extremely environment friendly. The mannequin makes use of roughly half the parameters of DeepSeek-R1 and one-third of Kimi-K2, but outperforms each in lots of benchmarks, indicating higher use of mannequin capability . Zhipu’s smaller 12B-active variant GLM-4.5-Air even surpassed sure bigger fashions like Google’s Gemini (2.5 “Flash” model), Alibaba’s Qwen-3 235B, and Anthropic’s Claude 4 Opus on some reasoning benchmarks . On a performance-per-parameter plot, the GLM-4.5 sequence lies on the Pareto frontier, that means it delivers best-in-class efficiency for its mannequin measurement . This emphasis on effectivity may enchantment to researchers and enterprises seeking to maximize AI functionality whereas managing compute prices.

Developments Over GLM-4 and Trade Friends

GLM-4.5 represents a major leap from Zhipu’s earlier era GLM-4 fashions. Whereas GLM-4 (launched earlier in 2024) supplied variations as much as 32B parameters and separate variants tuned for various duties  , the brand new GLM-4.5 consolidates a number of skillsets into one powerhouse mannequin. Zhipu calls it their “first full implementation” of merging various capabilities with out sacrificing any of them . In sensible phrases, GLM-4.5 can deal with duties that beforehand required a mixture of fashions – from complicated downside fixing and mathematical reasoning to writing code and controlling instruments – all inside a single unified system . This reduces the necessity for task-specific fashions (just like the prior GLM-4-Plus vs. GLM-4-Air variants) and simplifies deployment.

Due to the MoE design and optimization, GLM-4.5 achieves double the efficient mannequin capability of GLM-4 for a given computational price . The inference velocity and price enhancements are dramatic. Zhipu advertises GLM-4.5’s API at simply ¥0.8 CNY per million enter tokens and ¥2 per million output tokens, an order of magnitude cheaper than Claude’s API pricing . (For comparability, Claude 2’s 100k context API prices about ¥30 per million tokens.) This ~90% price discount lowers the barrier for companies to make the most of a GPT-4-class mannequin . Furthermore, a high-throughput model of GLM-4.5 can generate over 100 tokens per second, supporting low latency and high-concurrency functions . In contrast, GLM-4’s API had been priced larger (round ¥5 per million tokens) and will output only some dozen tokens/sec in commonplace settings  . GLM-4.5 thus delivers a serious improve in each efficiency and cost-efficiency over its predecessor.

When stacked towards Western friends, GLM-4.5 is aimed squarely at closing the hole. Zhipu’s launch messaging underscores a problem to OpenAI – the mannequin’s launch is timed as a direct bid to rival GPT-4 and the upcoming Google Gemini in functionality . Whereas GPT-4 stays proprietary, Zhipu claims GLM-4.5 attains comparable prowess on many benchmarks . The Chinese language press additionally highlighted that GLM-4.5’s parameter depend (355B) exceeds that of GPT-4, although as a consequence of MoE, its lively measurement is just like GPT-4’s estimated ~1.5T parameters scope (if every skilled counted) is bigger than GPT-4’s 175B, however as a consequence of MoE its lively measurement is ~32B, nearer to GPT-3.5/Claude – but it punches above its weight in efficiency  . Likewise, Anthropic’s Claude 2 (also referred to as Claude-4) has been lauded for its lengthy context and robust reasoning; GLM-4.5 matches Claude’s 100k token context with 128k, and Zhipu’s assessments point out GLM-4.5 can almost exchange Claude-4-Sonnet for coding assistant duties in enterprise settings . As for Google’s Gemini, nonetheless in improvement, Zhipu’s benchmarking suggests GLM-4.5 already competes: the smaller Air mannequin beating a “Gemini 2.5 Flash” baseline in some reasoning duties . These comparisons illustrate China’s accelerating progress – open fashions like GLM-4.5 are closing in on the capabilities of the very best closed fashions from US tech giants.

Enterprise Use Circumstances and Industrial Potential

By open-sourcing GLM-4.5 below the permissive MIT License, Zhipu has made it obtainable for business use by any developer or enterprise . Corporations can obtain the mannequin weights from platforms like Hugging Face and ModelScope, or put it to use by way of API on Zhipu’s cloud. This broad accessibility positions GLM-4.5 as a pretty basis mannequin for a wide range of real-world functions.

AI Brokers and Automation: GLM-4.5’s hallmark is enabling autonomous AI brokers. In the course of the launch, Zhipu showcased interactive demos of the mannequin performing as an AI agent throughout totally different situations. In a single demo, GLM-4.5 served as an online analysis assistant – given a subject, it may carry out stay searches, analyze the outcomes, and compile an aggregated reply with sources . One other demo featured a simulated social media platform the place the mannequin generated content material and even manipulated the interface (e.g. posting updates), demonstrating UI management talents . Impressively, the staff constructed a small Flappy Fowl recreation and had GLM-4.5 energy a bot to play the sport, displaying it dealing with front-end animation era and recreation logic management autonomously . Maybe essentially the most enterprise-relevant instance was an AI instrument for automated slide deck era: GLM-4.5 would take a subject, seek for related supplies and pictures, then dynamically generate a formatted PowerPoint in HTML – full with well-laid-out textual content and graphics  . The mannequin determined what data to incorporate and how one can model the slides, indicating potential for automating content material creation and information work.

Software program Growth: Given its sturdy coding abilities, GLM-4.5 can perform as a strong coding co-pilot or backend code era service. It helps code completion, debugging help, and multi-step instrument utilization (e.g. calling an API, then utilizing the lead to code). Zhipu experiences that GLM-4.5 carried out exceptionally in multi-turn coding duties that contain studying error messages, planning a repair, writing code, and verifying outputs . This makes it appropriate for constructing AI-assisted improvement instruments, from chatbots that assist engineers write code to totally automated coding brokers that may tackle outlined programming duties. The mannequin’s dependable instrument use (means to invoke exterior capabilities or APIs when wanted) and excessive job completion fee had been cited as key benefits on this area . Zhipu even made GLM-4.5 appropriate with Anthropic’s Claude Code framework, that means organizations which have workflows or agent methods designed for Claude can simply plug in GLM-4.5 instead AI engine . This interoperability lowers friction for enterprises to experiment with GLM-4.5 instead of current options.

Enterprise Intelligence and Lengthy Paperwork: With its 128k context window, GLM-4.5 can ingest and analyze very giant paperwork and even units of paperwork in a single go. Enterprises may leverage it for summarizing prolonged experiences, extracting insights from large logs or datasets, or conducting in-depth analysis by feeding the mannequin complete information bases. The mannequin additionally helps structured output codecs like JSON for simple integration into enterprise methods, and it encompasses a “information base retrieval” instrument interface to enhance its responses with factual references  . These options point out a give attention to enterprise wants comparable to accuracy, interpretability, and integration.

Value and Deployment: A significant promoting level is the low utilization price and versatile deployment choices. The official pricing of roughly ¥2.8 CNY per million tokens (enter+output mixed) is considerably cheaper than OpenAI or Anthropic’s APIs for comparable fashions . For firms that desire self-hosting as a consequence of information privateness or customization wants, GLM-4.5’s open supply nature permits deployment on non-public infrastructure (although the useful resource necessities for the 32B-active mannequin are non-trivial). The smaller GLM-4.5-Air lowers that bar and will probably run on extra inexpensive {hardware}. Moreover, Zhipu’s launch of high-speed inference optimizations (100+ tokens/sec era) means the mannequin can serve interactive functions to many customers concurrently with out main latency . All of those elements broaden the vary of enterprise use instances – from powering consumer-facing chat providers and digital assistants, to backend information evaluation pipelines – the place GLM-4.5 may very well be employed with manageable price and efficiency.

Strategic Place: China’s AI Ecosystem and World Competitors

Zhipu AI’s GLM-4.5 launch underscores China’s strategic embrace of open-source AI as a method to compete globally. By releasing a mannequin of this scale and class below an open license, Zhipu is “democratizing” superior AI capabilities  , in keeping with a broader pattern amongst Chinese language AI corporations. In current months, a number of Chinese language startups typically dubbed the “AI Six Tigers” (六小虎) have open-sourced giant fashions to quickly enhance and iterate. For instance, Beijing-based Moonshot AI open-sourced its Kimi K2 (a 1-trillion-parameter MoE mannequin) earlier in July, and Shenzhen’s Step AI (Jieyue Xingchen) launched an open model of its newest Step-3 reasoning mannequin  . Zhipu’s GLM-4.5, nevertheless, is the most important open mannequin thus far and positions the corporate as a front-runner on this open-source wave .

This technique is partly a response to U.S. gamers like OpenAI, whose cutting-edge fashions are closed-source and costly to entry. By providing a free (to obtain) various that approaches GPT-4-level efficiency, Chinese language firms hope to construct developer communities and set technical requirements within the AI enviornment  . The open-source strategy additionally serves as a “aggressive weapon,” permitting up-and-coming corporations like Zhipu to realize international traction regardless of much less assets than Large Tech incumbents . It encourages adoption by researchers and enterprises worldwide, probably tilting the community results in favor of those fashions over proprietary ones . As COAI Information notes, free high-quality fashions put stress on established gamers to rethink their pricing and entry insurance policies, and assist Chinese language corporations seize mindshare amongst builders who affect enterprise selections  .

Inside China, Zhipu is solidifying its function within the nationwide AI ecosystem. The corporate – a spin-off from Tsinghua College – was one of many first in China to develop a 100-billion-parameter mannequin (the GLM-130B in 2022) and has since been acknowledged as a number one innovator. It even drew consideration from OpenAI’s chief Sam Altman; Chinese language media famous that Zhipu was “named by OpenAI” as a notable Chinese language contender . Backed by main tech giants Alibaba and Tencent as buyers, in addition to funding from state-backed enterprise capital, Zhipu carries vital help  . Earlier this month, Shanghai’s Pudong VC and Zhangjiang Group invested ¥1 billion (~$140M) in Zhipu as a strategic stake . The corporate has raised a number of rounds from native governments in Hangzhou, Zhuhai, Chengdu and others in 2025 alone , aligning with China’s push to foster home AI champions. Zhipu is reportedly getting ready for an IPO, shifting its itemizing plan to Hong Kong with a possible ~$300 million elevate on the horizon . This infusion of capital would additional gasoline the compute and analysis wanted to compete with Western AI leaders.

Trade analysts see GLM-4.5’s launch as a pivotal second in China’s AI race. It comes on the heels of the World Synthetic Intelligence Convention 2025 (WAIC) in Shanghai, the place the divergence in methods amongst China’s AI startups was a scorching subject . Whereas a few of the “six tigers” have pivoted – e.g. Baichuan Clever specializing in medical AI, and ZeroOne dropping out of base mannequin R&D to focus on enterprise providers  – Zhipu has doubled down on the core basis mannequin strategy. By open-sourcing GLM-4.5, Zhipu goals to say technological management and guarantee its mannequin turns into a platform upon which others construct. This transfer additionally helps China’s nationwide aim of attaining AI self-reliance; open fashions may be extensively adopted domestically with out considerations of overseas restrictions. As one Chinese language commentary put it, open supply has turn into the “nook overtaking” technique for China’s AI – a option to leapfrog by harnessing group collaboration and transparency .

The worldwide AI group is taking notice as properly. GLM-4.5 being launched on Hugging Face and ModelScope means worldwide researchers can readily consider and fine-tune it . Its efficiency on English benchmarks (like MMLU and coding duties) suggests it’s aggressive not only for Chinese language-language functions however globally. If builders around the globe start adopting GLM-4.5 for novel functions, it may enhance China’s affect in AI software program akin to how open-source software program tasks have completed in different domains.

Launch Highlights and Outlook

Zhipu unveiled GLM-4.5 in a stay on-line occasion on the night of July 28, following WAIC. The launch included stay demonstrations of the mannequin’s agent capabilities (from coding to internet searching), which galvanized observers with their complexity . GLM-4.5 is straight away accessible: customers can check the total mannequin totally free by way of Zhipu’s internet portals, Zhipu Qingyan (chatglm.cn) and the z.ai platform, which host a chat interface . For builders, the mannequin’s API is accessible on Zhipu’s BigModel platform, and the entire educated weights may be downloaded from Hugging Face and ModelScope for experimentation . No particular business partnerships had been introduced on the launch, however Zhipu’s low-cost API and compatibility with current AI agent frameworks strongly trace at enterprise integrations within the close to future .

The response in China’s tech circles has been largely optimistic, with many seeing GLM-4.5 as an indication that native innovation is reaching a brand new stage. Some early Chinese language testers even described its efficiency as “离谱” (ridiculously sturdy) on sure duties . In fact, competitors is fierce and evolving: Baidu, Huawei, Alibaba, and others are all growing superior fashions (many open-sourced as properly) to compete in each home and worldwide markets.

For Zhipu, GLM-4.5 is a serious milestone but additionally a stepping stone. The corporate’s roadmap, supported by contemporary funding and an anticipated IPO, possible contains additional scaling of mannequin capabilities (GLM-5 and past) and specialization in domains like multimodal understanding (they lately open-sourced a vision-language mannequin, GLM-4.1V) . By staking a declare with an open mannequin that rivals the very best from Silicon Valley, Zhipu AI has firmly positioned itself on the vanguard of China’s AI development. GLM-4.5 not solely gives Chinese language enterprises an indigenous, cost-effective various for cutting-edge AI, but additionally alerts to the world that China’s open-source AI efforts are accelerating shortly. The worldwide AI group can count on extra to return as these fashions proceed to enhance and drive competitors – in the end benefiting customers and builders with extra selections and sooner innovation