血战太平洋
After the desert turns green, harder work begins_我的网站

一 | 连续多周包揽全球开源模型调用量排名前两席8月13日凌晨,杭州深度求索人工智能基础技术研究有限公司(DeepSeek),又悄悄搞了个大动作——DeepSeek V4 Pro正式版发布。此前,DeepSeek V4的另一个版本Flash正式版已于7月底发布。

二 |

Shi Jianyang (L) and his grandfather check the growth of crops on a farmland near a forest area along the Maowusu Desert in northwest China's Shaanxi Province in 2026.(Xinhua)A drone sweeps above the pine forest, sending images of the woods to a screen below. Under the trees stands Shi Jianyang whose grandfather and hundreds of villagers came 40 years ago with shovels and saplings to stop advancing sand.
At 29, Shi belongs to a new generation of desert-control workers equipped with drones, machines and remote irrigation systems. An earth auger can now do in a day what once required about 200 laborers, while a single phone can control drip irrigation across more than 1,000 mu (about 66.67 hectares) of woodland.
For the older Shi's generation, the challenge was getting trees to survive in moving dunes. For the grandson, it is figuring out how to make those forests healthier and more resilient for decades to come.
Their story captures a broader shift in desert control along the southern edge of the Maowusu Desert in northwest China's Shaanxi Province: turning shifting sand into forest was only the beginning.
In 1986, Shi Guangyin, Shi Jianyang's grandfather, led more than 300 villagers into Langwosha, a sandy area in Dingbian County in Yulin City. He had raised 60,000 yuan (about 17,377 U.S. dollars at the time) by selling his family's 84 sheep and borrowing from relatives and banks to contract 3,500 mu of barren sandy land.
The early setbacks were severe. Strong winds destroyed about 90 percent of the saplings in the first year, and about 80 percent of those planted the following year failed to survive.
After seeking advice from forestry experts, Shi and the villagers laid out checkerboard-shaped barriers on windward slopes to hold the sand in place, then planted trees, shrubs and grasses within them. Sapling survival rates rose to about 80 percent.
Over the next four decades, Shi and those working alongside him treated 250,000 mu of sandy and saline-alkali land, with more than 53 million trees and shrubs taking root.
Greening the Maowusu Desert is part of a much larger national undertaking. It is one of the three major battlefronts of China's Three-North Shelterbelt Forest Program, a decades-long effort to curb desertification across northern China.
China has made desertification control a national ecological priority. Since 2023, the country has stepped up efforts under the program, with greater emphasis on systematic restoration and on consolidating earlier gains.
Decades after the first trees took root, a new challenge emerged. Poplars planted decades ago for their rapid growth began reaching the end of their life cycle, with large numbers of trees dying, while forests dominated by a single species were more vulnerable to pests and disease.
Making a forest was one challenge. Keeping it healthy for another generation is another.
Shi Jianyang grew up watching his grandfather and father work to hold back the sand. His father, Shi Zhanjun, spent years planting trees alongside Shi Guangyin before dying in a road accident in 2008 while transporting saplings.
As a boy, Shi Jianyang sometimes resented the work that had taken his father. He spent summer holidays in makeshift tents in the sandy land, eating meals gritty with sand. During planting season, his father would return home with painful blisters in his mouth.
Yet he eventually chose forestry himself.
After studying forestry technology in university, Shi returned home in 2020 and began working in desert control. His team, initially made up of just seven or eight young locals, later brought in university graduates with expertise in soil science, landscaping and pest control.
The younger generation faced a different question: not simply how to make trees survive, but how to make a forest last.
Shi and his team spent more than 40 days in a 5,000-mu woodland setting up sample plots, measuring rainfall, studying soil microorganisms and tracking tree lodging. They proposed renewing aging forests with better-planned forest structures and expanded the planting of Mongolian pine, a long-lived evergreen species that had been successfully tested locally.
To date, about 50,000 mu of land has been planted with Mongolian pine. As some trees fail to survive and require replacement in subsequent planting seasons, the cumulative area of planting operations has reached about 65,000 mu.
Technology has also changed daily work. Drones patrol forests from the air, mechanized equipment dig planting pits, and irrigation can be managed remotely. Sapling survival rates, which Shi Guangyin's generation had raised to about 80 percent through improved sand-control and planting methods, have climbed further to about 92 percent with greater mechanization and more precise management.
Shi said the changes can be felt on the ground. Strong sandstorms have become rare, winds are weaker inside the forest, and the ground stays moist longer after rain. Summer days beneath the trees also feel cooler, while morning dew appears more often. The changes lead him to rethink desert control.
His approach also reflects a broader effort in China's desert regions to link ecological restoration with local livelihoods, giving communities a stronger economic stake in protecting restored land.
With his grandfather's support, Shi and his team have been looking for ways to make better use of the restored land as ecological conditions improve. They have developed potato seedlings suited to sandy soil, promoted greenhouse farming, and set up a livestock cooperative, while increasingly growing potatoes, watermelons and peppers on land once prone to losing water and nutrients.
Shi has also turned to livestreaming e-commerce to market potatoes and other local farm produce, while developing forest-based health and leisure projects. More than 2,000 local residents have seen their incomes rise steadily as farming, livestock breeding and other businesses have developed.
"We used to fear the sand and keep away from it. Now we understand it and know how to make use of it," Shi Guangyin said. "Today, there are opportunities to make a living everywhere in this sandy land."
Across Yulin, forest coverage has risen from 1.8 percent in 1949 to 34.8 percent, while shifting dunes that once encroached on villages have been brought under control.
For Shi Jianyang, however, the green landscape is not proof that the work is finished. "Planting a tree is not difficult," he said. "The difficult part is looking after it for decades."
The grandfather generation learned how to stop the sand. The young generation is learning how to make what grew there last.

Shi Jianyang (R) and his grandfather check tree growth in a forest area in Dingbian County in Yulin City, northwest China's Shaanxi Province in 2026. (Xinhua)
。从官方测评数据看,Pro版本能力大幅提升,多项指标已接近Anthropic Fable 5等国外顶尖闭源模型。

三 | 而在十天前,另一家浙江公司阿里巴巴发布旗舰模型Qwen3.8-Max,首次亮相即进入权威评测平台LMArena综合榜Top 5。全球开源模型调用量排名中,两个浙产大模型DeepSeek与千问已连续多周包揽前两席。人工智能正成为创新浙江最鲜明的标识、最核心的支撑。而选择“开源”来加速技术迭代、构建产业生态,成为浙产大模型企业一条坚定而清晰的路径。这是企业对技术路径和商业模式的选择,也是浙江发展人工智能的战略共识。正如中国工程院院士、之江实验室主任王坚反复提及的观点:人工智能开源,正从开放源代码即“Open Source”,转为开放资源即“Open Resource”。中国开放AI模型,本质上意味着把背后的资源和机遇开放给全世界。性能做强,门槛降低“便宜的没它强,强的没它好用。

四 | ”8月13日,杭州嗡嗡科技有限公司负责人刘杲劼试用了DeepSeek V4 Pro版本。他告诉记者,此前公司刚从国外模型切换到了7月底发布的DeepSeek V4 Flash版本,它在编程等工作上表现不错,费用降低了不少。这是DeepSeek和千问广受欢迎的原因:够便宜,够好用。便宜到什么程度?DeepSeek V4 Flash最近采用了分时定价机制,在空闲时段,每百万输出Token仅需4.5元人民币,即便在高峰时段,价格也仅为9元/百万输出Token。相比顶尖的国外模型,价格优势依然显著,两者在代码编写能力相差不大,官方测试数据显示仅差几个百分点。

五 | 好用到什么程度?DeepSeek V4 Pro总参数1.6万亿、每次激活约490亿参数,V4 Flash总参数2840亿、激活130亿参数,两者全系开源,均支持百万Token超长上下文。Qwen3.8-Max同样如此:总参数量达2.4万亿,单次推理激活950亿参数,采用稀疏混合专家架构,支持最大1M Token的上下文窗口。这意味着开发者更方便在其上自由搭建、改造,并商业化落地。杭州电子科技大学计算机学院副院长余宙告诉记者,国内的开源大模型集中精力做了两件事——把模型性能做强,把使用门槛降低。“中小企业很难从0到1训练一个大模型。”浙江古珀医疗科技有限公司是杭州一家专注医疗大数据领域的科技企业,创始人张强说,对开发者而言,这两大开源模型提供了成本更低、性价比更高的选择,拉低了人工智能现实落地的门槛。古珀就利用开源模型Qwen3.6-35B-A3B和Qwen3.6-27B,打造了一个面向C端的医疗智能体,具有“医生分身+智能随访+慢病风险评估智能体+健康管理”等功能。截至目前,浙江累计已完成90款生成式人工智能服务备案。其中,有一些模型,就是利用千问、DeepSeek等基础大模型进行二次开发和参数调整获得的。融入全球,改写规则两个“顶流”大模型的背后,是浙江在战略层面锚定的路线。浙江省“十五五”规划纲要中六次提及“开源”,频次之高、着墨之重,前所未有。

六 | 2026年省政府工作报告提出,“积极抢占人工智能发展制高点,着力营造最优开源开放生态”,将开源列为全省AI布局的核心抓手之一。7月16日,浙江省委十五届九次全会在杭州举行。全会期间进行了人工智能专题辅导授课,组织考察了国家人工智能应用中试基地(具身智能)、“魔搭社区”(杭州)开发者中心、杭州城西科创大走廊未来科技城展馆。其中,“魔搭社区”是国内最大的人工智能开源社区。魔搭社区以“模型即服务”的模式,线上汇聚全球超2900万人工智能开发者、超20万个模型,极大降低了人工智能应用创新的门槛。浙江大学人工智能学院教授、本科生院院长吴飞解释了开源的底层逻辑:“开源是开放知识与创造能力。

七 | AI技术具有先行者引领的溢出效应,需要去主动构建开源生态。”“极致性价比+开放生态”的浙产大模型,正在重写全球AI竞争的游戏规则。7月24日,AI硬件巨头英伟达创始人黄仁勋,发布了一封公开信呼吁开放模型权重。该信件将开放模型比作1980年代开源软件运动——开放权重扩大了参与AI经济的途径。

八 | 初创企业、成熟企业、高校及公共机构无需从零开始训练模型,也不必为每一项任务都支付高昂的尖端模型使用费,而是可以直接基于现有的先进模型进行开发。日前,联合国人工智能问题独立国际科学小组发布报告《对人工智能带来的机遇、风险和影响的询证评估》。

九 | 该报告将Qwen及DeepSeek作为全球开源AI生态代表性案例单独列出。

十 | 这既是技术层面的认可,也是国际社会对中国开源治理理念的一种肯定。
Current article:http://h4scq5.yepenchunshouzuiwai.shop/eofw/20260826/6133908.html
Published on:07:25:10