不少企业用户出于隐私和保密，对于本地大模型具有刚性需求，这种趋势被硬件厂商和中小AI团队视为新的商业机遇。
新华财经北京9月14日电（记者叶健） 随着腾讯、阿里、字节等行业巨头发力办公智能体赛道，另一条基于本地大模型的赛道也开始兴起。业内人士认为，随着云端大模型不断抬高智能上限，本地大模型的智能水平也持续提升，结合智能体正加速融入办公场景。不少企业用户出于隐私和保密，对于本地大模型具有刚性需求，这种趋势被硬件厂商和中小AI团队视为新的商业机遇。
本地大模型进入办公场景
9月初，北京大学科学智能学院孵化的元空AI发布自研本地大模型Boxer，以及基于此的AI办公平台和AI科研平台。记者在发布会现场看到，在一台40TOPS（每秒万亿次运算）算力的消费级PC上，对话延迟接近真人交流，运行内存占用约8GB。
据悉，Boxer本地大模型基于阿里千问的Qwen3.5-35B-A3B后训练而来。元空智能联合创始人宁鲲鹏表示，团队在2023年就发布了AI产品ChatExcel，已积累数百万用户。
经过“后训练”强化的中小参数大模型正加速进入办公场景。近日，原点星辉发布本地大模型StartLux-V1.0-27B-Preview，它以阿里千问Qwen3.6-27B为基座模型，面向工具调用与通用智能体场景进行后训练加强。在工信部中国信息通信研究院可信AI大模型基准测试的MCP（模型上下文协议）专项测试中，StartLux综合性能超越多款千亿参数大模型。
原点星辉CTO郭权玮表示，当AI工作得越来越深，隐私、成本和持续运行就会从附加条件变为真正影响体验的核心问题。本地不再只是一个部署选择，而会成为下一代智能越来越自然的形态。
小模型打开办公新空间
与云端大模型相比，更小的模型为何能够承担更多的办公任务？
一方面，企业越发注重AI安全。元空智能创始人逄大嵬认为，企业级AI的竞争不在于模型大小，在于数据是否出设备、成本、使用稳定性，而这正是本地大模型和端侧AI的产业机会。
中国惠普副总裁、个人信息产品事业部总经理周信宏告诉记者，对于企业用户而言，使用云端大模型的最大风险是企业数据、经验被“蒸馏”。
另一方面，本地大模型的智能化水平正在快速崛起。郭权玮告诉记者，云端大模型在继续推高智能上限，本地大模型则在享受三条技术曲线的红利——大模型本身越来越高效、消费级硬件越来越强、推理和智能体系统越来越成熟。
逄大嵬也认为，从生产力与成本平衡维度评估，20B至100B参数的本地大模型是当前最适合规模化交付的区间，可支撑长链路智能体任务，如今35B大模型的智能化水平已经可以媲美前两年的一线大模型。
据悉，目前能运行20B至100B大模型的电脑，价格区间在1万元至6万元，这也被很多厂商视为AI硬件的新机遇。
周信宏告诉记者，受内存、CPU等核心部件涨价影响，目前低价PC市场持续萎缩。惠普在大中华区持续发力AI端侧领域布局，PC营收保持两位数增长，其核心驱动力正是产品单价持续提升。
从“能用”到“好用”仍需闯关
尽管供需双方都很看好本地大模型及其应用，但“办公”场景更看重如何“干活”。
商业与技术洞察公司Gartner管理副总裁孙鑫认为，中小参数模型虽然降低了本地推理门槛，但企业采用的门槛仍在于数据与应用执行。一个智能体能否按权限找到正确版本的文件、调用业务系统、在失败后恢复，比能装下多少参数更能决定其办公价值。此外，评估成本也要包含硬件、维护和人工返工。
本地大模型无需“包打天下”。郭权玮表示，在应用场景方面，更看好那些高频、长期，以及用户数据与工具深度绑定的工作场景。比如，长期整理电脑中的文件和邮件、梳理企业内部知识和流程、开展科研资料分析。这些任务真正需要的不是偶尔调用一次模型，而是AI能否一直助力工作。
在AI巨头发力通用办公市场的同时，专业办公场景也值得关注。逄大嵬表示，公司将优先深耕低容错率的垂直行业场景，例如，精密制造、医疗、科研、财务等场景，持续沉淀高价值行业能力与数据，由此构建高替换成本的产品壁垒。
编辑：罗浩
本文来源：https://www.cnfin.com/gs-lb/detail/20260914/4469262_1.html
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Many corporate users have rigid demand for local large models due to privacy and confidentiality, which hardware manufacturers and small-to-medium AI teams view as a new commercial opportunity.
Beijing, September 14 (Xinhua Finance, Reporter Ye Jian) — As industry giants like Tencent, Alibaba, and ByteDance exert strength on the office intelligent agent track, another track based on local large models is also emerging. Industry insiders believe that as cloud-based large models continuously raise the upper limit of intelligence, the intelligence level of local large models is also improving, accelerating integration into office scenarios through combination with intelligent agents. Many corporate users have rigid demand for local large models due to privacy and confidentiality, which hardware manufacturers and small-to-medium AI teams view as a new commercial opportunity.
Local large models enter office scenarios
In early September, Yuan Kong AI, incubated by Peking University’s School of Intelligent Science, released the self-developed local large model Boxer, as well as the AI office platform and AI research platform based on it. The reporter observed at the press conference that on a consumer-grade PC with 40 TOPS (tera operations per second) computing power, the conversation delay is close to human interaction, and the running memory usage is approximately 8 GB.
It is understood that the Boxer local large model was post-trained based on Alibaba’s Thousand Questions Qwen3.5-35B-A3B. Ning Kunpeng, co-founder of Yuan Kong Intelligent, stated that the team had released the AI product ChatExcel in 2023, already accumulating millions of users.
Small-to-medium parameter large models that have been “post-trained” are accelerating entry into office scenarios. Recently, Origin Starshine released the local large model StartLux-V1.0-27B-Preview, which uses Alibaba’s Qwen3.6-27B as the base model and has been post-trained for tool calling and general intelligent agent scenarios. In the Ministry of Industry and Information Technology’s China Academy of Information and Communications Trusted AI Large Model Benchmark Test’s MCP (Model Context Protocol) special test, StartLux’s comprehensive performance surpassed that of multiple trillion-parameter large models.
Guo Quanwei, CTO of Origin Starshine, said: “As AI works more and more deeply, privacy, cost, and continuous operation will become core issues that truly affect the experience rather than additional conditions. Local will be more than just a deployment choice; it will become the increasingly natural form of the next generation of intelligence.”
Small models open new spaces for office work
Compared with cloud-based large models, why can smaller models undertake more office tasks?
On one hand, enterprises are increasingly focused on AI security. Pang Dawei, founder of Yuan Kong Intelligent, believes that the competition in enterprise-level AI does not lie in the size of the model but in whether data leaves the device, cost, and stability of use — these are the industrial opportunities for local large models and edge AI.
Zhou Xinhong, vice president of HP China and general manager of the personal information product division, told the reporter: “For corporate users, the biggest risk of using cloud-based large models is that enterprise data and experience are ‘distilled’.”
On the other hand, the intelligence level of local large models is rapidly rising. Guo Quanwei said that while cloud-based large models are continuing to raise the upper limit of intelligence, local large models are enjoying the dividends of three technical curves: the large model itself is becoming increasingly efficient, consumer-grade hardware is becoming more powerful, and the inference and intelligent agent system is maturing.
Pang Dawei also believes that from the perspective of balancing productivity and cost, local large models with 20B to 100B parameters are the most suitable range for large-scale delivery at this stage, capable of supporting long-chain intelligent agent tasks. At present, the intelligence level of a 35B large model is already comparable to that of leading large models from two years ago.
It is understood that computers that can currently run 20B to 100B large models are priced between 10,000 and 60,000 yuan, which many manufacturers also view as a new opportunity for AI hardware.
Zhou Xinhong said that due to the price increase of core components such as memory and CPU, the low-price PC market is currently shrinking continuously. HP has continued to strengthen its layout in the edge AI field in Greater China, with PC revenue maintaining double-digit growth — the core driving force is the continuous increase in product unit price.
Still need to overcome challenges to go from “usable” to “easy to use”
Although both supply and demand sides are optimistic about local large models and their applications, the “office” scenario relies more on how to “get the job done.”
Sun Xin, vice president of Gartner, a business and technology insight company, believes: “Although small-parameter models lower the threshold for local inference, the threshold for enterprise adoption still lies in data and application execution. Whether an intelligent agent can find the correct version of a file according to permissions, call business systems, and recover after a failure is more critical to its office value than how many parameters it can accommodate. Additionally, evaluation costs must include hardware, maintenance, and manual rework.”
Local large models do not need to “do it all”. Guo Quanwei said he is more optimistic about those work scenarios that are high-frequency, long-term, and where user data and tools are deeply bound — such as long-term organization of files and emails on the computer, sorting internal knowledge and processes of enterprises, and carrying out scientific research data analysis. These tasks do not require occasional model calls but rather whether AI can continuously assist in work.
While AI giants are exerting strength in the general office market, professional office scenarios are also worth noting. Pang Dawei said the company will prioritize deep cultivation of vertical industry scenarios with low error tolerance, such as precision manufacturing, medicine, scientific research, and finance, to continuously accumulate high-value industry capabilities and data, thereby constructing a product barrier with high replacement cost.
编辑：罗浩