Retail News CRM

Tag: zhipu ai

  • China Mobile Cloud Signs Zhipu to Link GLM Models and Build Coding Tools

    China Mobile Cloud and Zhipu signed a cooperation pact on November 6 to connect the GLM model family to the operator’s MoMA service platform.

    The agreement establishes a joint development center focused on token generation efficiency, targeting domestic hardware adaptations alongside financial services and government contracts.

    Technical teams from both companies plan to tune Zhipu’s models specifically for Chinese-made computing processors. The partnership outlines shared research into inference performance, yet neither company published baseline efficiency figures or an operational schedule for customer deployments. Access arrangements, service-level commitments and supported versions within the GLM lineup were also omitted from the joint filing.

    Hardware Adaptation and Coding Tools

    Beyond platform hosting, the two companies plan to develop joint AI coding applications for commercial software engineers. The announcement outlined an integrated product program without releasing software names, development roadmaps or public testing dates.

    For enterprise buyers, the lack of a timetable leaves procurement teams waiting on functional specifications. Developers evaluating enterprise assistants must weigh unreleased software against tested tools already running on commercial clouds across mainland China.

    Rival Hyperscalers Build Token Tariffs

    Commercial cloud operators in China are locking enterprise accounts into aggressive subscription packages rather than unpriced development agreements. Alibaba Cloud already offers its AI Token Plan starting at $6 a month on its Model Studio service, bundling its proprietary Qwen lineup alongside DeepSeek and GLM models. Alibaba also sells dedicated enterprise engineering assistants, such as Qoder, directly to corporate IT departments.

    State-backed telecom operators often hold strong connections to public sector budgets, but commercial developers demand transparent pricing and tested inference pipelines. China Mobile Cloud must show that domestic silicon can match commercial cloud economics on cost per million tokens. Without published benchmarks, enterprise procurement officers will treat the venture as an exploratory project rather than an immediate supplier switch.

    From Transport Compute to Platform Hosting

    China Mobile has been steering its infrastructure divisions toward specialized vertical computing workloads. The carrier previously launched an automotive AI foundation featuring 112 exaflops of computing power and six specialized agents for vehicle manufacturers. Adding Zhipu’s models to MoMA extends that infrastructure strategy into general business software, banking risk engines and municipal administration.

    Commercial validation now hinges on the release of concrete access tiers, supported GLM versions and billing schedules on China Mobile Cloud’s MoMA catalog.

  • Huawei Cloud Retires 2 openPangu Models from Model-as-a-Service Platform

    Huawei Cloud Retires 2 openPangu Models from Model-as-a-Service Platform

    Huawei Cloud will shut down commercial access to its openPangu-2.0-Flash and openPangu-2.0-Pro AI models on its Model-as-a-Service platform, ending sales on November 6 and terminating all hosted calls on November 30.

    The cut comes four months after Huawei open-sourced the 505-billion-parameter Pro version, forcing enterprise customers running consumer analytics and internal search workloads in its Southwest Guiyang cloud region to switch tools before December.

    New customers cannot buy or activate either model service after midnight on November 6, while existing accounts retain operational access only through the end of the month. Huawei Cloud directs accounts using openPangu-2.0-Flash to DeepSeek-V4.1-Flash-1017 as their replacement, while routing openPangu-2.0-Pro users to Zhipu AI’s GLM-5.3.

    Migration Demands and Technical Friction

    Engineering teams face immediate rewrite costs. Huawei Cloud noted that the replacement models carry different API parameters, cost structures and output profiles, requiring developers to retest task pipelines rather than swap endpoints blindly.

    Software setups built on the company’s early MaaS standard API V1 must be rewritten to work with an OpenAI-compatible endpoint. Teams already on API V2 or an OpenAI-style interface need to alter model identifier calls.

    Retail commerce platforms relying on openPangu for live inventory categorisation or chatbot assistance now have less than a month to validate replacement models. Changing baseline models often alters latency profiles and customer response formats across live digital storefronts.

    Rival Platforms and Vendor Strategy

    The decision to retire homegrown models while recommending independent rivals highlights the fierce pace of platform rationalisation across China’s cloud ecosystem. Rather than carrying the operational cost of maintaining earlier foundation models, cloud providers now drop their own releases rapidly in favour of top-performing external open weights.

    DeepSeek and Zhipu AI gain immediate enterprise volume from the shift. Huawei Cloud’s public embrace of DeepSeek-V4.1-Flash and GLM-5.3 confirms that third-party open architectures are setting operational benchmarks on Chinese sovereign server hardware, even inside Huawei’s proprietary Ascend computing clusters.

    Platform developers bear the stability risk. By cycling through model versions within single-digit months, platform operators transfer the cost of code maintenance, integration testing and deployment downtime onto enterprise software clients.

    Rapid Model Turnover on Ascend Hardware

    Huawei made openPangu-2.0-Flash public on June 30 with 92 billion parameters, followed by the 505-billion-parameter openPangu-2.0-Pro sparse model on July 31. Both variants were tailored for Ascend neural processing units, accompanied in late September by the release of complete pretraining, supervised fine-tuning and reinforcement learning source code.

    The company’s hosted platform has cleared out earlier systems at a comparable pace throughout the autumn. Huawei Cloud scheduled GLM-5.1 for removal on November 16, switched DeepSeek-V4-Flash accounts to version 4.1, set Kimi-K2.6 for retirement on November 8, and pulled DeepSeek-V4-Pro from hosting in September alongside older Qwen and Wan models.

    The retirement schedule applies strictly to managed API services hosted out of the Guiyang data centre, leaving the underlying open-source weights and training code available in public repositories for teams running their own compute stacks.

    Software teams on Huawei Cloud must finish transitioning their active workloads to the DeepSeek and GLM services before the Guiyang cluster unplugs both openPangu endpoints at midnight on November 30, while Huawei prepares a successor generation of Pangu models without a public release date.

  • Zhipu AI Launches GLM-5.3-Flash After Stealth Trial on 100,000 Domestic Chips

    Zhipu AI Launches GLM-5.3-Flash After Stealth Trial on 100,000 Domestic Chips

    Beijing-based Zhipu AI launched its open-weight model GLM-5.3-Flash on Wednesday after running the system across a cluster of 100,000 domestic Chinese chips. The release followed an unannounced trial on platforms including OpenRouter and OpenCode, where the system processed 62 trillion tokens under the test alias Ox Alpha before its formal unveiling.

    Shares tied to the startup rose following the disclosure. Zhipu confirmed that the entire test deployment operated on domestic silicon rather than foreign hardware imports.

    Traffic Across Developer Marketplaces

    The stealth evaluation generated heavy traffic across international developer hubs over the past week. Users tested the then-unidentified Ox Alpha across code generation and complex task routing, driving sustained compute volume across the 100,000-chip array before Zhipu claimed ownership of the checkpoint.

    Releasing the architecture as open weights allows enterprise clients and independent software developers to download, modify, and host the model on their own infrastructure. That deployment model lowers operating expenses for commercial software firms looking to integrate natural language tools without paying continuous per-token API charges to proprietary providers.

    Domestic Compute Strategy

    Building massive cluster capacity on domestic processors addresses direct supply constraints that Chinese technology groups face under ongoing hardware export restrictions. Large tech firms across the mainland have traditionally relied on foreign graphics processors to train top-tier foundation models, but local alternatives are now taking on larger workloads.

    The performance of the 100,000-unit setup provides an operating template for other Chinese software developers seeking to decouple their deployment pipelines from foreign accelerators. Market attention now turns to downstream enterprise adoption rates and independent performance benchmarks against competing commercial models over the current quarter.

  • Chinese AI and Chip Makers Hand Millions in Stock to Lock Down Staff

    Chinese AI and Chip Makers Hand Millions in Stock to Lock Down Staff

    Chinese artificial intelligence and semiconductor firms are granting stock awards worth millions of dollars to staff across entire workforces to halt talent poaching.

    Chip designer Cambricon Technologies unlocked nearly 600,000 shares for 124 core staff, delivering an average payout of 5.57 million yuan (US$828,000) per employee. The Shanghai-listed company also launched a 5 million share scheme covering 944 workers through 2028, representing 85.3 per cent of its total payroll.

    Hardware Giants Expand Workforce Coverage

    Zhongji InnoLight, which makes optical transceivers for AI data centres, allocated 2.48 million shares across 99 key personnel in its latest vesting cycle. Based on April market pricing, that distribution yielded an average of more than 26 million yuan per recipient.

    Equipment manufacturer Advanced Micro-Fabrication Equipment China widened access further. Its March restricted share programme enrolled more than 97 per cent of its workforce. GPU maker Moore Threads followed in April by assigning 1,080 employees, roughly 85 per cent of its headcount, into its equity pool. At memory producer ChangXin Memory Technologies, founder Zhu Yiming pledged 768 million shares, half his partnership stake, to fund a decade-long employee pool while excluding himself from payouts.

    Zero Targets and Pre-IPO Payouts

    AI model developers are structuring payouts with fewer performance hurdles. Hong Kong-listed MiniMax issued 1.16 million zero-cost shares in June to core personnel, tying vesting schedules solely to tenure rather than operational benchmarks. Rival developer Zhipu AI allocated a 9.8 per cent post-listing stake across 426 staff through an internal platform, yielding an average holding value exceeding HK$100 million per person.

    Big Tech platforms are stepping up their own programmes to match startup offers. Tencent Holdings granted more than 38.6 million shares under its employee scheme, representing 0.42 per cent of its issued equity, as Alibaba Group Holding, Baidu, Meituan and Xiaomi rework compensation packages.

    The scale of these handouts reflects how quickly technical headcount costs have risen across mainland China. Where earlier stock plans favoured senior directors, current programmes distribute equity down to floor engineers to insulate operations from overseas recruitment drives and venture-backed rivals.

    Market watchers are now monitoring upcoming interim financial filings to track the share dilution costs from these multi-year vesting programmes.