文 | 数字力场，作者 | 佘宗明
在国产大模型沉寂两个多月引发“国模已经不行了吗”质疑声四起的背景下，Kimi出来放招了。
10月24日凌晨，K3.1 Pro正式登场。
如果说K3代表了上个版本的“开源模型天花板”，那K3.1 Pro就是百尺竿头更进N步的K3进化版。
能不能打，其表现自己会说话。
Artificial Analysis将其智能指数排在了全球第四，还给出判断：一旦权重发布，K3.1 Pro将登顶开放权重模型榜首，顶级开源模型与闭源模型的差距将从12分降至3分。
在行业公认硬核的终端Agent编码基准Terminal-Bench4.0上，K3.1 Pro的成绩是65.7分，仅次于Cluade Opus5.5的66.4分，开源模型和闭源模型的差距进一步缩小。
置于当前形势中看，这实属来之不易。
01
要知道，过去2个多月，是国模的暗淡期。
大洋彼岸，分外热闹：9月份，OpenAI 先后发布了GPT-6 Astra和GPT-6.1 Sol，A社捧出了Claude Opus 5.5，次月，就连此前被认为掉队了的谷歌，都凭着Gemini 4 Argon重上牌桌……
虽然都在“必须给前沿AI踩刹车”的呼吁上摆足了姿态，可不得不说，御三家身体很诚实，你方唱罢我登场，轮流坐庄闭源SOTA，高频的上新节奏让硅谷AI圈恨不能捡起“人类的黄金时代”梗。
国模这边，则俨然与之形成了鲜明对照：自Kimi K3引爆全球科技圈后，国模就有些沉寂。模型厂商推出新模型的动作有是有，但激起的声量几乎可以忽略不计。
要而言之：人们盼“牛来”，可牛未来。
如此一来，多个权威榜单上的Top10大模型，近乎成了美国AI模型的内部点名册。
不怕沉寂，就怕对比。伴随强烈反差而来的，是“蒸馏指控”与“能力怀疑”的甚嚣尘上。
“蒸馏”向来是Anthropic们站在道德制高点上扔向国模的飞镖，这两个月尤甚。从NSA+CISA+FBI三大机构出手，到Anthropic又双叒叕打小报告，再到OpenAI首次公开发布反蒸馏行动报告，都来得极具针对性。
而国模的沉寂发生在Anthropic们强化反蒸馏防御机制后，很容易被视作“变相坐实”。
在Kimi K3将顶尖开源模式跟顶级闭源模型的技术代差大幅抹平，把“时间差”从1年以上缩小到约4个月后，这两个月的反差，又让许多人对“开源模型可以赶超闭源模型”的判断产生了怀疑，觉得差距又被拉开了。
更有甚者，断定开源模型突破靠的就是蒸馏，反蒸馏无异于抽走了开源模型往上攀爬的梯子。
犹记得，K3刚问世时，我写了篇《取次硅谷懒回顾，半缘DeepSeek半缘Kimi》。可不得不说，近3个月的情况，让我生出了“不会被打脸了吧”的担忧。
在此背景下，K3.1 Pro的“救场”，来得正是时候。它无形中释放了信号：越过灰暗时刻，China is back。
02
“蒸馏论”的默认逻辑是：蒸馏可以帮你加速追赶，但不能帮你成为Sota——蒸馏是跟跑捷径，但不是领跑路径。
月之暗面有关负责人此前曾做过回应：K3 性能跃升的核心来自底层原创架构创新，并非对现有模型蒸馏复刻。而K3突破依赖的那些关键创新点——Moon Clip 二阶优化器+KDA混合线性注意力+Attention Residuals 注意力残差+Stable LatentMoE 稀疏专家框架，明显都不是蒸馏出来的。
此次K3.1 Pro在Anthropic们开启“反蒸馏大法”的前提下仍实现了升级，就给“国模全靠蒸馏论”提供了最直观的反例。
在算力与资金不占优的掣肘下，国模跟硅谷前沿模型的差距接下来是会拉大还是会缩小，也是很多人关心的问题。
得看到，自7月下旬以来，国内模型厂商发布的模型都是Flash版（高速普惠版），没有Pro版（旗舰版），跟御三家连发旗舰版比，势能上就弱了些。K 3.1 Pro则填补了国模这几个的“Pro版空窗”。
这里没有说Pro版跟Flash版有好坏之分的意思，但从性能维度看，二者确有强弱之别：一般来说，Pro对应的是“能力上限优先”路线；Flash对应的是“效率/成本优先”路线。
在其他厂商纷纷推出Flash版的情况下，Kimi依旧用Pro版去摸上限，这不免又让人想Cue起梁静茹老师。
K3.1 Pro相当于把开源模型见“先”思齐的势头又给续上了。考虑到跟硅谷模型厂商的算力资源约束条件大不一样，能把中美顶尖模型智能指数差距收窄到3分，真的不容易。
更何况，K3.1 Pro在审美这块也很在线，一如既往地很显Taste。
03
从K3到K3.1 Pro，Kimi证明了自身的技术迭代是可持续的。
从技术创新角度看，做出一次惊艳的突破和建立一个稳定的突破机制，是两码事。
前者可能是灵光乍现、运气加成的产物；后者则意味着企业找到了从灵感、实验、工程到产品的完整链路，这条链路能经受起时间的检验。
若Kimi推出K3后后继乏力，那K3也就是座孤峰，可“连续版本号”K3.1 Pro的出现，表明K3连着的是山脉。
事实上，Kimi创新迭代的连续性早已有迹可循——
K2出来时，就被黄仁勋点赞、马斯克评论，还被《自然》网站视作“又一个DeepSeek时刻”；
K2.5出来后，被Cursor“致敬”，Cursor创始人公开称K2.5是其测评中“最强的基座模型”，马斯克也盖章“Yeah，it’s Kimi 2.5”；
之后的K2.6首个实现的大规模并行Agent集群+浏览器常驻后台Worker，也在程序员社区备受认可……
这种连续性对模型厂商很重要。时下的很多模型厂商，创新曲线是脉冲式的：憋个大招，刷屏一时，然后……就没有然后了。
这不免让人想起电影《浪浪山小妖怪》里的剧情：小猪妖获得了父亲传下来的一个大招，可惜一辈子只能施展一次。
中国AI，要的显然不是一次性惊艳，而是连续性进化。
着眼长远看，AGI就是个以十年为尺度的技术长征。
爱跑马拉松的跑者都知道，跑马拉松不能前半程冲太猛、后半程泄了气，真正能完赛的，是那些配速稳定、耐力持久的人。
AGI征程也一样。既是长跑，长久韧性比短暂锐度更重要。
04
说起月之暗面，许多人都会想起，这四个字是对摇滚乐队Pink Floyd同名专辑的致敬。
细想来，“月之暗面”本就有“一半是光明，一半是未知”的寓意。将这份寓意放在Kimi身上，也很贴切：它的来路上满是光明，它的前路上遍布未知。
现在看，马拉松远未结束。但很显然，Kimi正在领跑之列。
李小龙曾在笔记中写道：截拳道，使我们一旦确定了方向，即不再回首反顾。
明确方向后一直向前走，很重要。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, digital force field, author, Zhongming
Kimi came out of his mouth against the backdrop of the questioning that the national production of LLM had been silent for more than two months, “Is the national model gone?”
In the early hours of 24 October, K3.1 Pro officially arrived.
If K3 represents the last version of the Open Source Model ceiling, then K3.1 Pro is the K3 evolutionary version of 100-foot poles moving up the N step.
Can't fight, it shows that you can talk.
Artificial Analysis ranks its intelligence index at the fourth highest in the world and also makes the judgement that the gap between the top open source model and the closed source model will be reduced from 12 to 3 points once the weight publish, K3.1 Pro, will top the top open weight model.
K3.1 Pro achieved 65.7 points on the industry-recognized hard-core terminal Agent code benchmark Terminal-Bench 4.0, followed by 66.4 points on Clube Opus 5.5, further reducing the gap between open-source and closed-source models.
It was not easy to put it in the current situation.
01
You know, it's been more than two months, a bleak period for the national model.
On the other side of the ocean, it's fun: in September, Open AI, after publish GPT-6 Astra and GPT-6.1 Sol, A, with Claude Opus 5.5, next month, even Google, which was previously considered to be out of line, reset the table with Gemini 4 Argon...
Despite all the gestures in the call for “the need to stop the front-line AI”, it must be said that the three families are honest, that you are going to stop me, that you are going to sit in the closed-source SOTA, and that the new rhythm of high frequency has prevented Silicon Valley Ai's hatred from picking up “the golden age of humanity”.
This side of the national model is in stark contrast: it has been a bit silent since Kimi K3 detonated the global scientific circle. There are moves by model manufacturers to introduce new models, but the amount of noise generated is almost negligible.
In other words, people are looking forward to "Cows".
As a result, Top10LLM, which is on several authoritative lists, has almost become the internal point roster of the United States AI model.
And fear not silence, but contrast. This is accompanied by a strong contrast between “still allegations” and “capability doubts”.
The "distillation" has always been the darts of the Anthropics who were thrown at the high moral point to the national model, especially in the last two months. From NSA+CISA+FBI to Anthropic's double-crossing reports, to Open AI's first release of the publish anti-distillation report, it's very specific.
And the silence of the national model, which took place after the Anthropics reinforced the anti-distillation defense mechanism, could easily be regarded as “concealed”.
The two-month contrast between Kimi K3 and Kimi K3 had significantly eliminated the technological gaps between the top open-source model and the top closed-source model, narrowing the “time gap” from more than one year to about four months later, causing many to question the judgement that “open-source models can catch up with closed-source models” and felt that the gap had been closed.
Moreover, it is concluded that the open source model is broken by distillation, which is tantamount to removing the ladder from the open source model.
Remember, when K3 first came into existence, I wrote a book entitled " Taking Silicon Valley Lazy Review, DeepSeek Semi-Kimi ". It must be said that the situation in the last three months has given me the fear of not being hit in the face.
In this context, K3.1 Pro's “Suspense Field” came at the right time. It's an invisible signal: past the gray hour, China is back.
02
The default logic of distillation theory is that distillation can help you catch up, but it can't help you become Sota -- distillation is the shortcut to running, but not the lead path.
The person responsible for the dark side of the moon had previously responded: the core of the K3 leap was from the bottom-up original architecture innovation, not the retort of existing models. And the key innovation points on which the K3 breakthrough depends - the Moon Clip 2 optimizer + KDA mixed linear attention + Attention Resicuals Attention Disorders + Stable LatemoE Rare Expert Framework – are obviously not distilled.
K3.1 Pro was upgraded this time with the Anthropics’ introduction of the “reverse distillation technique”, providing the most intuitive reverse to the “national model based on distillation theory”.
The gap between the National Model and the Silicon Valley frontier model, under pressure from poor calculation and finance, will be widened or narrowed, and is of concern to many.
It must be seen that, since late July, the models of the domestic model manufacturer publish have all been in the Flash version (high-speed inclusive version), without the Pro version (flag version), and have been weaker compared to the three flagship versions. K 3.1 Pro filled the "Pro version" windows of the National Model.
It is not said that Pro and Flash mean good or bad, but there is a difference in performance: In general, Pro corresponds to the “capability ceiling priority” route; Flash corresponds to the “efficiency/cost priority” route.
Kimi continues to use the Pro to touch the ceiling in the context of the launch of the Flash version by other manufacturers, which in turn gives rise to the idea that Cue will raise Teacher Liang.
K3.1 Pro is equivalent to a renewed momentum to see the open source model “first” in mind. Given the great differences in the calculator resource constraints of Silicon Valley models, it is not easy to narrow the smart index gap between the top Central American models to three points.
Moreover, K3.1 Pro is also online in the aesthetic review, showing Taste as always.
03
From K3 to K3.1 Pro, Kimi demonstrated that his technology is sustainable.
From the perspective of technological innovation, it is two things to make an amazing breakthrough and to establish a stable breakthrough mechanism.
The former may be the product of light and luck; the latter means that the enterprise has found a complete link from inspiration, experiment, engineering to product, which can be tested over time.
If Kimi's K3 is weak, then K3 is an isolated peak, but the emergence of the "continuous version" K3.1 Pro suggests that K3 is connected to mountains.
In fact, the continuity of Kimi's innovation has long existed -
When K2 came out, he was praised by Hoang In-hoon, Mask, and viewed by Nature as “another DeepSeek moment”;
When K2.5 came out, it was “honored” by Cursor, whose founders publicly called K2.5 “the strongest base model” in their assessment, and Mask also stamped “Yeah,it's Kimi 2.5”;
The first large-scale parallel Agent Cluster + Browser to be realized in K2.6 was also recognized in the programmer community at Worker's backstage.
This continuity is important for model manufacturers. Many of the current model manufacturers, the innovation curve is pulsive: hold a big move, brush a screen, and then... there is no more.
It reminds me of the story in the film "The Little Wanderer": Piggy demons have a great move from their father, but only once in their lives.
China AI, obviously, does not want a one-time surprise, but a continuous evolution.
In the long run, AGI is a decade-scale technological journey.
Those who run the marathon know that the marathon can't run the first half too hard, the second half of the race is really finished by those who have a steady, durable end.
The same goes for AGI. It's both long run and long-term resilience is more important than short-lived sharpness.
04
When it comes to the dark side of the moon, many will remember that the four words are a tribute to the same name as the rock band Pink Floyd.
To think about it, the dark side of the moon was meant to be half bright and half unknown. It's also very appropriate to put it on Kimi: Its journey is full of light, and its path is full of unknowns.
Now, the marathon is far from over. But apparently Kimi is in the lead.
In his notes, Bruce Lee wrote: Cutting through boxing makes us look back once we have set our course.
It is important to move forward with a clear direction.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.