本文来源:https://so.html5.qq.com/page/real/search_news?docid=70000021_0446afd11ac13752
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
2.8 万亿参数、全面开源、国产芯片完成训练适配——月之暗面把开源旗舰的天花板又抬高了一截。
11 月 18 日消息,月之暗面正式开源旗舰模型 Kimi K3.1 Pro,总参数规模达 2.8 万亿,是目前开源权重模型里的第一梯队体量。更值得划重点的是硬件适配:摩尔线程已完成 K3.1 Pro 的训练适配,这意味着这个万亿级模型从第一天起就验证了「国产 GPU 能养得起国产开源旗舰」——开源不只是放权重,算力底座同样在去依赖。
开源旗舰的意义要放在整个行业的开源-闭源对峙里看。过去半年,闭源阵营接连亮出超大参数旗舰,开源社区则用「够用+自由」守住开发者基本盘。K3.1 Pro 把 2.8 万亿参数的权重直接放出来,等于告诉所有开发者:你可以自己部署、自己微调、自己蒸馏,不用被任何 API 定价和限流政策拿捏。对企业用户来说,开源旗舰意味着私有化部署的选项——金融、医疗、政务这些数据不出域的行业,终于能在顶级模型能力上做选择,而不是在「安全」和「能力」之间二选一。
时间线上还有个耐人寻味的巧合:阿里刚发布 2.4 万亿参数的 Qwen4-Max 且首次不开放权重,月之暗面转头就把 2.8 万亿参数的开源旗舰放出来。一个是「闭源收紧」,一个是「开源加倍」,两条路线在同一个星期正面相撞。对开发者来说这是最好的时代——两边卷得越凶,能白嫖到的能力上限就越高。
摩尔线程的适配是一条被低估的信息。过去国产 GPU 的处境有点尴尬:软件生态跟不上,模型厂商适配意愿弱,用户不敢上生产环境。而 K3.1 Pro 这种万亿级模型的训练适配,等于给国产 GPU 做了一次最高强度的压力测试——能跑通这个量级的训练,推理和微调场景基本就是顺手的事。开源模型当「适配标准」,国产算力当「量产底座」,这个组合跑顺了,国内 AI 基础设施的自主化就形成了闭环。
从商业模式看,月之暗面的打法也在进化:旗舰开源换生态位,API 和企业服务收商业化的钱。这个路径已经被验证过——开源权重带来的社区口碑和下游衍生模型,反哺的是商业版的调用量。对开发者来说还有一个现实好处:大参数开源模型是蒸馏的优质「教师」,中小团队可以从它蒸馏出适合自己业务的小模型,一次开源,养活整条下游产业链。
对整个国产大模型阵营来说,这次开源还有一层防御价值:当海外闭源模型因为地缘因素存在服务不确定性时,一个能力在线、权重在手、国产算力可跑的旗舰模型,就是国内应用开发者的「保底选项」。保底选项平时不一定用,但它的存在本身就是安全感——这种安全感的定价,远不止一个模型权重。
2.8 万亿参数的旗舰不是给个人显卡准备的——满血版需要多卡集群,个人玩家更多是通过官方 API 或托管平台体验。真正值得动手的是三类人:一是做私有化部署的企业工程师,重点测长上下文和文档处理能力,这是 Kimi 系的传统强项;二是做模型蒸馏的算法团队,权重开放就是原料库;三是做国产算力方案集成的服务商,K3.1 Pro+摩尔线程的组合可能就是明年政企客户的标配套餐。开源模型战争进入「万亿俱乐部」时代,你更看好开源还是闭源?评论区站个队。
本文来源:https://so.html5.qq.com/page/real/search_news?docid=70000021_0446afd11ac13752
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
2.8 trillion parameters, fully open-sourced, and training adaptation completed on domestic chips - Moonshot AI has raised the ceiling for open-source flagships again.
On November 18, Moonshot AI officially open-sourced its flagship model Kimi K3.1 Pro, with 2.8 trillion total parameters, placing it in the top tier of open-weight models. More important is the hardware adaptation: Moore Threads has completed training adaptation for K3.1 Pro, meaning this trillion-scale model proves from day one that domestic GPUs can support a domestic open-source flagship. Open source is not just releasing weights; the compute base is also being de-dependent.
The significance of an open-source flagship must be seen in the industry-wide open-versus-closed confrontation. Over the past six months the closed camp has unveiled one ultra-large flagship after another, while the open-source community has held the developer base with capable and free models. Releasing the 2.8-trillion-parameter weights amounts to telling every developer: you can deploy, fine-tune, and distill it yourself, without being held hostage by API pricing or rate limits. For enterprises, an open flagship means a private-deployment option - data-sensitive sectors such as finance, healthcare, and government can finally choose top-tier model capability instead of picking between security and capability.
There is an intriguing coincidence in the timeline: Alibaba had just released the 2.4-trillion-parameter Qwen4-Max without opening weights for the first time, and Moonshot immediately released a 2.8-trillion-parameter open flagship. One tightens closed source, the other doubles down on open source; the two routes collided head-on in the same week. For developers this is the best of times - the harder they compete, the higher the capability ceiling you can get for free.
Moore Threads' adaptation is an underrated piece of news. Domestic GPUs were in an awkward position: software ecosystems lagged, model makers were reluctant to adapt, and users dared not deploy to production. Training adaptation for a trillion-scale model like K3.1 Pro is a maximum-intensity stress test for domestic GPUs - if training at this scale works, inference and fine-tuning are essentially routine. With open models as the adaptation standard and domestic compute as the production base, this combination, once smooth, closes the loop of self-reliance in domestic AI infrastructure.
Commercially, Moonshot's approach is also evolving: open-source the flagship to win ecosystem position, and monetize through APIs and enterprise services. This path has been validated - community reputation and downstream derivative models from open weights feed back into commercial call volume. For developers there is another practical benefit: a large open model is a high-quality teacher for distillation, letting small and mid-sized teams distill smaller models suited to their business.
For the whole domestic large-model camp, this release also has defensive value: when overseas closed models carry service uncertainty for geopolitical reasons, a flagship with online capability, weights in hand, and runnable on domestic compute is the fallback option for domestic application developers.
A 2.8-trillion-parameter flagship is not meant for personal graphics cards - the full version needs a multi-GPU cluster, and individual users mostly experience it through official APIs or hosted platforms. Three groups are truly worth acting: enterprise engineers doing private deployment, focusing on long context and document processing (a traditional Kimi strength); algorithm teams doing model distillation, for whom open weights are the raw material; and service providers integrating domestic compute solutions, for whom K3.1 Pro plus Moore Threads could be the standard package for government and enterprise clients next year. The open-model war has entered the trillion-parameter club era.
