One other week in the summertime of 2025 has begun, and in a continuation of the pattern from final week, with it arrives extra highly effective Chinese language open supply AI fashions.
Little-known (at the very least to us right here within the West) Chinese language startup Z.ai has launched two new open supply LLMs — GLM-4.5 and GLM-4.5-Air — casting them as go-to options for AI reasoning, agentic conduct, and coding.
And based on Z.ai’s weblog put up, the fashions carry out close to the highest of the pack of different proprietary LLM leaders within the U.S.
For instance, the flagship GLM-4.5 matches or outperforms main proprietary fashions like Claude 4 Sonnet, Claude 4 Opus, and Gemini 2.5 Professional on evaluations comparable to BrowseComp, AIME24, and SWE-bench Verified, whereas rating third general throughout a dozen aggressive checks.
Its lighter-weight sibling, GLM-4.5-Air, additionally performs throughout the prime six, providing sturdy outcomes relative to its smaller scale.
Each fashions function twin operation modes: a pondering mode for complicated reasoning and gear use, and a non-thinking mode for fast response situations. They will automatically generate full PowerPoint shows from a single title or immediate, making them helpful for assembly preparation, schooling, and inside reporting.
They additional provide inventive writing, emotionally conscious copywriting, and script technology to create branded content material for social media and the online. Furthermore, z.ai says they help digital character growth and turn-based dialogue programs for buyer help, roleplaying, fan engagement, or digital persona storytelling.
Whereas each fashions help reasoning, coding, and agentic capabilities, GLM-4.5-Air is designed for groups looking for a lighter-weight, extra cost-efficient different with sooner inference and decrease useful resource necessities.
Z.ai additionally lists a number of specialised fashions within the GLM-4.5 household on its API, together with GLM-4.5-X and GLM-4.5-AirX for ultra-fast inference, and GLM-4.5-Flash, a free variant optimized for coding and reasoning duties.
They’re obtainable now to make use of instantly on Z.ai and thru the Z.ai utility programming interface (API) for builders to connect with third-party apps, and their code is out there on HuggingFace and ModelScope. The corporate additionally offers a number of integration routes, together with help for inference through vLLM and SGLang.
Licensing and API pricing
GLM-4.5 and GLM-4.5-Air are launched below the Apache 2.0 license, a permissive and commercially pleasant open-source license.
This permits builders and organizations to freely use, modify, self-host, fine-tune, and redistribute the fashions for each analysis and industrial functions.
For individuals who don’t need to obtain the mannequin code or weights and self-host or deploy on their very own, z.ai’s cloud-based API presents the mannequin for the next costs.
- GLM-4.5:
- $0.60 / $2.20 per 1 million enter/output tokens
- GLM-4.5-Air:
- $0.20 / $1.10 per 1M enter/output tokens
A CNBC article on the fashions reported that z.ai would cost solely $0.11 / $0.28 per million enter/output tokens, which can be supported by a Chinese language graphic the corporate posted on its API documentation for the “Air mannequin.”
Nonetheless, this seems to be the case just for inputting as much as 32,000 tokens and outputting 200 tokens at a single time. (Recall tokens are the numerical designations the LLM makes use of to symbolize totally different semantic ideas and phrase parts, the LLM’s native language, with every token translating to a phrase or portion of a phrase).
In truth, the Chinese language graphic reveals much more detailed pricing for each fashions per batches of tokens inputted/outputted. I’ve tried to translate it under:
One other notice: since z.ai relies in China, these within the West who’re targeted on information sovereignty will need to due diligence by way of inside insurance policies to pursue utilizing the API, as it might be topic to Chinese language content material restrictions.
Aggressive efficiency on third-party benchmarks, approaching that of main closed/proprietary LLMs
GLM-4.5 ranks third throughout 12 business benchmarks measuring agentic, reasoning, and coding efficiency—trailing solely OpenAI’s GPT-4 and xAI’s Grok 4. GLM-4.5-Air, its extra compact sibling, lands in sixth place.
In agentic evaluations, GLM-4.5 matches Claude 4 Sonnet in efficiency and exceeds Claude 4 Opus in web-based duties. It achieves a 26.4% accuracy on the BrowseComp benchmark, in comparison with Claude 4 Opus’s 18.8%. Within the reasoning class, it scores competitively on duties comparable to MATH 500 (98.2%), AIME24 (91.0%), and GPQA (79.1%).
For coding, GLM-4.5 posts a 64.2% success charge on SWE-bench Verified and 37.5% on Terminal-Bench. In pairwise comparisons, it outperforms Qwen3-Coder with an 80.8% win charge and beats Kimi K2 in 53.9% of duties. Its agentic coding capacity is enhanced by integration with instruments like Claude Code, Roo Code, and CodeGeex.
The mannequin additionally leads in tool-calling reliability, with a hit charge of 90.6%, edging out Claude 4 Sonnet and the new-ish Kimi K2.
A part of the wave of open supply Chinese language LLMs
The discharge of GLM-4.5 arrives amid a surge of aggressive open-source mannequin launches in China, most notably from Alibaba’s Qwen Workforce.
Within the span of a single week, Qwen launched 4 new open-source LLMs, together with the reasoning-focused Qwen3-235B-A22B-Pondering-2507, which now tops or matches main fashions comparable to OpenAI’s o4-mini and Google’s Gemini 2.5 Professional on reasoning benchmarks like AIME25, LiveCodeBench, and GPQA.
This week, Alibaba continued the pattern with the discharge of Wan 2.2, a strong new open supply video mannequin.
Alibaba’s new fashions are, like z.ai, licensed below Apache 2.0, permitting industrial utilization, self-hosting, and integration into proprietary programs.
The broad availability and permissive licensing of Alibaba’s choices and Chinese language startup Moonshot earlier than it with its Kimi K2 mannequin displays an ongoing strategic effort by Chinese language AI firms to place open-source infrastructure as a viable different to closed U.S.-based fashions.
It additionally locations strain on the U.S.-based mannequin supplier efforts to compete in open supply. Meta has been on a hiring spree after its Llama 4 mannequin household debuted earlier this 12 months to a blended response from the AI group, together with a hearty dose of criticism for what some AI energy customers noticed as benchmark gaming and inconsistent efficiency.
In the meantime, OpenAI co-founder and CEO Sam Altman not too long ago introduced that OpenAI’s long-awaited and much-hyped frontier open supply LLM — its first since earlier than ChatGPT launched in late 2022 — could be delayed from its initially deliberate July launch to an as-yet unspecified later date.
Structure and coaching classes revealed
GLM-4.5 is constructed with 355 billion whole and 32 billion lively parameters. Its counterpart, GLM-4.5-Air, presents a lighter-weight design at 106 billion whole and 12 billion lively parameters.
Each use a Combination-of-Consultants (MoE) structure, optimized with loss-free stability routing, sigmoid gating, and elevated depth for enhanced reasoning.
The self-attention block contains Grouped-Question Consideration and a better variety of consideration heads. A Multi-Token Prediction (MTP) layer permits speculative decoding throughout inference.
Pre-training spans 22 trillion tokens break up between general-purpose and code/reasoning corpora. Mid-training provides 1.1 trillion tokens from repo-level code information, artificial reasoning inputs, and long-context/agentic sources.
Z.ai’s post-training course of for GLM-4.5 relied upon a reinforcement studying part powered by its in-house RL infrastructure, slime, which separates information technology and mannequin coaching processes to optimize throughput on agentic duties.
Among the many strategies they used have been mixed-precision rollouts and adaptive curriculum studying.
The previous assist the mannequin prepare sooner and extra effectively by utilizing lower-precision math when producing information, with out sacrificing a lot accuracy.
In the meantime, adaptive curriculum studying means the mannequin begins with simpler duties and progressively strikes to tougher ones, serving to it be taught extra complicated duties progressively over time.
GLM-4.5’s structure prioritizes computational effectivity. In line with CNBC, Z.ai CEO Zhang Peng acknowledged that the mannequin runs on simply eight Nvidia H20 GPUs — customized silicon designed for the Chinese language market to adjust to U.S. export controls. That’s roughly half the {hardware} requirement of DeepSeek’s comparable fashions.
Interactive demos
Z.ai highlights full-stack growth, slide creation, and interactive artifact technology as demonstration areas on its weblog put up.
Examples embody a Flappy Chook clone, Pokémon Pokédex net app, and slide decks constructed from structured paperwork or net queries.
Customers can work together with these options on the Z.ai chat platform or by way of API integration.
Firm background and market place
Z.ai was based in 2019 below the title Zhipu, and has since grown into one in every of China’s most distinguished AI startups, based on CNBC.
The corporate has raised over $1.5 billion from traders together with Alibaba, Tencent, Qiming Enterprise Companions, and municipal funds from Hangzhou and Chengdu, with further backing from Aramco-linked Prosperity7 Ventures.
Its GLM-4.5 launch coincides with the World Synthetic Intelligence Convention in Shanghai, the place a number of Chinese language corporations showcased developments. Z.ai was additionally named in a June OpenAI report highlighting Chinese language progress in AI, and has since been added to a U.S. entity listing limiting enterprise with American corporations.
What it means for enterprise technical decision-makers
For senior AI engineers, information engineers, and AI orchestration leads tasked with constructing, deploying, or scaling language fashions in manufacturing, the GLM-4.5 household’s launch below the Apache 2.0 license presents a significant shift in choices.
The mannequin presents efficiency that rivals prime proprietary programs throughout reasoning, coding, and agentic benchmarks — but comes with full weight entry, industrial utilization rights, and versatile deployment paths, together with cloud, non-public, or on-prem environments.
For these managing LLM lifecycles — whether or not main mannequin fine-tuning, orchestrating multi-stage pipelines, or integrating fashions with inside instruments — GLM-4.5 and GLM-4.5-Air scale back boundaries to testing and scaling.
The fashions help customary OpenAI-style interfaces and tool-calling codecs, making it simpler to judge in sandboxed environments or drop into present agent frameworks.
GLM-4.5 additionally helps streaming output, context caching, and structured JSON responses, enabling smoother integration with enterprise programs and real-time interfaces. For groups constructing autonomous instruments, its deep pondering mode offers extra exact management over multi-step reasoning conduct.
For groups below funds constraints or these looking for to keep away from vendor lock-in, the pricing construction undercuts main proprietary options like DeepSeek and Kimi K2. This issues for organizations the place utilization quantity, long-context duties, or information sensitivity make open deployment a strategic necessity.
For professionals in AI infrastructure and orchestration, comparable to these implementing CI/CD pipelines, monitoring fashions in manufacturing, or managing GPU clusters, GLM-4.5’s help for vLLM, SGLang, and mixed-precision inference aligns with present greatest practices in environment friendly, scalable mannequin serving. Mixed with open-source RL infrastructure (slime) and a modular coaching stack, the mannequin’s design presents flexibility for tuning or extending in domain-specific environments.
In brief, GLM-4.5’s launch provides enterprise groups a viable, high-performing basis mannequin they will management, adapt, and scale, with out being tied to proprietary APIs or pricing buildings. It’s a compelling possibility for groups balancing innovation, efficiency, and operational constraints.
