文 | 象先志
用上K3.1 Pro，我的第一反应是：快了。哪怕用的是Pro，也比之前的K3快了很多。
K3刚出来的时候，速度确实难以评价。这次测试的时候，感觉一切都丝滑无比。虽然肯定还达不到各家flash那种感觉，但是完全是正常偏快的思考速度。
之前一堆人喊着蒸馏被ban之后国产大模型就不做pro了，Kimi这不就直接出来打小黑子的脸了。
速度思考能力又开始结合
这次我没有开高速模式，普通Pro的速度就已经让我满意。Kimi还推出了Pro Highspeed，按官方说法，效果与普通Pro相同，模型输出速度约为它的2—3倍。
但这不是K3到K3.1的提速幅度。普通Pro为什么变快，官方没有单独解释推理端做了哪些调整。技术博客更多谈的是训练：Kimi增加了后训练投入，让模型在数百个下游领域中接受联合强化学习，也改进了长上下文处理和执行任务的环境。
榜单里，执行任务的成绩涨得最明显。独立评测机构Artificial Analysis公布的综合指数中，K3.1 Pro从K3的44分升到55分，排名第4。终端任务测试Terminal-Bench 4.0则从12.6%升到67.2%，排在该次公布结果的首位。它要求模型在电脑终端中完成任务，通过全部检查才算成功。
对于非技术背景的使用者，办公相关的两项成绩更有参考价值。GDPval-AA比较模型制作文档、表格和演示文稿的质量，K3.1 Pro排到第3。AutomationBench-AA考的是跨邮箱、表格和客户管理系统完成工作的能力，Pro的目标完成得分从58.3%升到了73.3%。
在Kimi公布的成绩表中，Pro也拿到了两项领先：长任务编程测试SWE-Marathon v1.1为51.7%，高于Opus 5.5的46.3%；财务资料问答OfficeQA Pro v2为64.4%，高于Opus的62.2%。两项均排在该表所列模型的首位。这里Pro的成绩由Kimi自测，对手成绩部分来自外部榜单。
进步并不均匀。在官方HLE全量难题测试中，使用工具的Pro拿到70.0%，高于Opus 5.5；不用工具时，Pro只有51.1%，低于Opus的59.8%。AA的长上下文推理测试里，K3（Max）原有88.7%，Pro是86.3%。
官方的人体行走演示更直观：一个虚拟人起初不断摔倒，经过几版训练，逐渐能连续行走。K3.1 Pro编写模拟器、安排训练和评估，再根据实验结果修改方法。过程中，手臂动作一直不理想，调整奖励也没有解决；模型随后检查模拟器，修复了36处肩部肌肉力臂计算错误，再重新训练控制器。
人体的动作由仿真和控制策略产生，Kimi负责根据结果继续改进训练。它没有停在调整奖励上，还找到了模拟器里的错误，这部分比最后走得多像人更让我感兴趣。
做这类长任务，调用量自然会增加。Kimi这次把普通Pro的API输出价格降了下来：每百万输出token，K3为100元，Pro为68元，Highspeed则要136元。Token是平台计算文字和代码用量的单位，包月用户仍按套餐额度使用。
Kimi Code的高速版也要消耗约两倍额度，而且提速只发生在模型输出阶段。读写文件、执行命令和运行程序，仍然需要原来的时间。普通Pro已经够快时，我会先把额度留给后面的修改。
速度上来了，额度未必够用
有用户觉得普通Pro够快了。他贴出的一条调用记录达到每秒111个token，表示不开高速版也够用。
获得Kimi官方提前访问的Chetaslua则放出了网页演示，自述是模型一次生成，并附上CodePen入口，读者可以直接打开看效果。
也有用户在来回修改时碰到了额度问题。
LanceYang给了模型五张参考图，让它做一个可交互的3D游戏手柄。第一版之后又改了两轮，部件形状和空间关系还是不对，模型验证修改结果时又碰到了用量限制。他用的是199元套餐，开始做手柄之前已经用掉了一部分额度。
在知名的鹈鹕骑单车测试中，另一个帖子里的两位用户报出了很不同的消耗。有用户用199元套餐，选高速版和最高思考档，生成后又要求“继续优化”，滚动额度就接近用完。Sanyeg用599元套餐，模型档位相同，却说一个鹈鹕任务只消耗了约6%。套餐、任务和修改过程都不同，很难只看比例判断哪个更划算。所以现在整体看来，token消耗略微有点迷。
按Kimi的规则，各项功能共享月度额度，模型、上下文和工具调用都会影响消耗，Code另外还有滚动用量限制。用户能否顺利完成一个作品，要看从生成到修改的全部用量。
场景有小bug，但能自修复
我选了七道编程题，每道都新建会话，允许Kimi自行运行、检查和修改，结束后再验收首次交付。我们没有替它改代码。图片、动图和提示词节选放在下面，完整提示词及输入素材附在发布包中。
“鹈鹕骑自行车”已经被测得太多，怕有题库优化。所以我们选择：让自行车骑鹈鹕
实际提示词（节选）：画面左右两个同等大小的舞台同时播放，左边表现“鹈鹕骑自行车”，右边表现“自行车骑鹈鹕”。……自行车由鹈鹕背部承托，两个车轮都离开地面，鹈鹕双腿交替迈步并向前行进。
Kimi做了左右对照：左边鹈鹕蹬车，右边鹈鹕走路，自行车悬在它上方。位置换过来了，车和鸟背之间却空着一截，看不出鹈鹕在承托自行车。这部分应该是逻辑上鹈鹕和自行车前面的节点连接，所以导致后面的悬空，另外一个小问题是自行车的车轮还在一直转，这点逻辑上有点问题，但是也是可以进行优化的。
物理题的表现更好。我们设置了小球在转动的七边形里碰撞，观察实际的碰撞和运行规律，效果还是很不错的。
实际提示词（节选）：正七边形外接圆半径320，以0.2弧度/秒顺时针转动。内部放20个半径12的小球，编号1至20、初始互不重叠，受竖直向下500像素/秒²的重力；……墙面反弹需考虑旋转墙面的局部速度。
通过持续运行模拟、检查小球位置，没有发现穿墙或丢球。它也把墙面的运动速度算进了反弹过程。这道题交出的初稿，已经可以用来演示重力和碰撞。
声音分析题，我另外换了低音和高音文件，检查频谱是否真能反映输入。
实际提示词（节选）：通过Web Audio API分析实际播放的声音，显示频谱、RMS响度和主频Hz；频率轴范围至少0至4000Hz，使用可解释的真实数据，禁止预编节奏动画。音量滑块影响输出和分析结果。……附带F05-demo.wav可用于开发，但程序须支持其它同格式音频。
换成低音和高音文件后，图表与频率读数都能对应。音量减半，测出的振幅也接近减半；静音后，振幅归零，主频数值不再显示。
不过，短促的声音会在图上拖一会儿才消失。用它看声音的高低、强弱已经够用，想让画面紧跟鼓点，还得调整。
像素跑者只有12×8像素，要用四帧动作表现奔跑。
实际提示词（节选）：用最多32种固定RGB颜色画七彩道路、星空和12×8像素的跑者。跑者有可辨识的四帧跑动循环，身体始终保持12×8逻辑尺寸；道路向左滚动，角色停留在画面水平中心。
尺寸、颜色数量和动作帧数都符合要求。小人的胳膊和腿交替摆动，七彩道路向后滚，画面虽然简单，跑动的动作已经能辨认出来。
还有一道，让同一张卡片在按钮、播放器等控件之间连续变形。
实际提示词（节选）：画面中心同一个卡片外壳在16秒内依次变成按钮、加载器、成功勾、播放器、音量控件、开关、柱状图、提示条，再回到按钮。……用形状、圆角、尺寸、图标和文字的连续变换表现状态，不能仅用整页切换。
大部分变化都顺畅，问题出在最后：提示文字先消失，画面留下空白按钮，下一圈开始时，按钮上的文字突然出现。看截图容易忽略，连续播放就能发现。
在测评3D建模能力的部分，我们选择了清明上河图的案例，并且特意要求人物和船都动起来。过桥时，脚下的高度要跟着桥面变化；搬运工则要先取货，再走到码头放下。把镜头拉近，才看得清它有没有把动作和场景配合好。
实际提示词（节选）：请制作一个可在浏览器中自由漫游的3D《清明上河图》场景，重点表现汴河、木拱虹桥、桥头集市、沿河商铺和远处城门。……人物要有可辨认的身体和肢体动作，持续进行不同活动，至少包括沿街行走、过桥、摊前交易，以及在码头搬运货物；船要沿河航行，船上人物随船运动。……行人沿可通行道路、码头和桥面移动，脚贴着当前地面或桥面高度，不走进河里、建筑、摊位，也不从栏杆穿过去。
制作过程中，我看到小人离地，也看到有人走进河里、穿过铺子。这些问题最初都会发生。然后Kimi还是自行优化。
Kimi开始自己逐步调整，穿模问题、人物走过河的问题，这些基本的逻辑问题我们并没有额外增加提示词，并且它还自行修正了人物的高度，也改了穿过商铺的行走路线。
最牛的是，阴影部分的计算都是非常准确的，可以明显感觉模型已经开始理解物理世界的真实规律了。
这些修改发生在Kimi自己的检查过程中，我们没有追加修改提示。
交付版里，搬运工会拿起货物走到码头，再空手返回，顾客也会在摊前停下来。近看仍能发现问题：行人走动时，身体会穿进商铺墙面。河道也有问题：水面被地面挡住，船看起来像在地上滑行。
最后，我让Kimi做了一套离线视频剪辑器。时间线、预览和导出得使用同一份编辑结果，刷新网页也得保留刚才的修改。
实际提示词（节选）：每个片段可以精确设置入点、出点；支持在当前播放头位置分割片段、删除片段、重新排列片段顺序，以及撤销和重做编辑。……真正导出可播放的 WebM 视频，允许实时导出。导出必须反映剪切和重排后的内容，不能下载原始视频冒充导出。……本地保存项目状态；刷新或重新打开后，能恢复已导入的素材与编辑结果。
我们把跑者和鹈鹕两段录像导进去，各截取一段，调整成先跑者、后鹈鹕。检查导出的视频，顺序和截取范围都对应得上。刷新网页，素材和剪过的时间线也恢复了。分割、删除、撤销和重做在这轮操作里都正常，这套初稿已经能完成这次简单剪辑。
导出文件还有个小毛病：在测试用的Edge浏览器里，刚打开时读不到视频总时长，要播完一次才能识别。视频内容正确，文件信息还没处理完整。
这一轮里，物理、音频分析和简单视频剪辑已经有实际用途。几道视觉题也交出了可以继续修改的初稿，空间关系和动效衔接仍需检查。
Kimi把Pro称为K3.1系列的首个模型，目前还没有公布Flash或Ultra。如果以后推出Flash，我更看重它能不能降低频繁调用的成本。
多个Agent同时工作，会产生更多的模型调用。如果Flash能以更低的价格、更快的速度处理简单步骤，开发者就能把复杂问题留给Pro。
此前有报道，月之暗面开始招募具身智能人才，一些机器人公司的员工接到了Kimi的招聘电话。报道还援引参与面试人士的说法，称内部对Agentic Robotics（智能体机器人）路线仍有保留。
Google已经尝试把Flash用于机器人。它以Gemini 3.5 Flash为基础，经过专门的具身推理训练，做出了Gemini Robotics ER 2。这个模型负责观察环境、规划和监测任务，再调用下层动作模型执行。
人体行走的演示，再加上具身人才招聘，很难让人不相信，Kimi也想在这个领域继续探索。
我目前最看好的具身方案，就是让强模型参与设计和改进训练，让控制策略在仿真中学会动作；实际运行时，由快速模型观察环境、规划任务，再交给下层控制模型执行。训练可以反复试，机器人面对变化中的环境时，却要及时作出反应。假如Kimi未来有低延迟、低成本的多模态Flash，它就可能用在观察和规划这一部分。
至于Ultra，我期待的是它补上更难的推理和长任务能力。Pro在不用工具的难题测试中仍有差距。
虽然这次测试没有“震惊瘫坐”，但其实可以推断出Kimi在朝着真实物理世界方向以及整体的能力优化提升有着比较稳步的推进节奏，毕竟这次还是在K3的底座上进行优化。
Kimi已经交卷了，现在开源就等Qwen4了。
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。
Wen Wei Wei Wei Wei Wei Wei
With K3.1 Pro, my first reaction is: soon. Even with Pro, it's a lot faster than K3.
When K3 came out, the speed was indeed difficult to evaluate. This test feels like everything's silky. It certainly does not reach the sense of the individual flash, but it is entirely normal to think fast.
Before, a bunch of people shouted that after the distillation of Ban, the LLM, the national product, would not be pro, and Kimi would just come out and hit Blackie in the face.
The ability to think fast begins to merge again.
I didn't open the high-speed mode this time, and I'm satisfied with normal Pro speed. Kimi also introduced Pro Highspeed, which, in official terms, has the same effect as normal Pro, with model output rates of about two to three times its rate.
This, however, is not the rate of K3 to K3.1. The ordinary Pro has not been able to explain separately what adjustments have been made to the reasoning. Technical bloggers talk more about training: Kimi has increased his post-training input, allowing the model to undergo joint intensive learning in hundreds of downstream areas, and has improved the long context processing and mission environment.
In the list, performance in the implementation of the mandate has increased most significantly. Of the composite index published by the independent evaluation body Artificial Analysis, K3.1 Pro rose from 44 to 55 points in K3, ranked fourth. The end-task test Terminal-Bench 4.0 rose from 12.6 per cent to 67.2 per cent, ranking first in the publication of the results. It requires that the model be completed in a computer terminal, and only successful through a full check.
For users with non-technical backgrounds, two of the office-related achievements are more relevant. GDPval-AAA compares the quality of documentation, tables and presentations, K3.1 Pro to 3. Automation Bench-AAA examined the ability of cross-mails, forms and customer management systems to complete their work, and Pro's target achievement score increased from 58.3 per cent to 73.3 per cent.
In the accomplishment form published by Kimi, Pro also received two lead: the long mission programming test SWE-Marathon v1.1 was 51.7 per cent, up from 46.3 per cent in Opus 5.5; and the financial information QA Pro v2 was 64.4 per cent, up from 62.2 per cent in Opus. Both are at the top of the models listed in the table. Here, Pro's scores are measured by Kimi, and the scores of rivals are partly from external lists.
Progress is uneven. In the official HLE full-volume challenge test, the tool-using Pro received 70.0 per cent, higher than Opus 5.5; when the tool was not used, Pro was only 51.1 per cent below 59.8 per cent of Opus. In the long context reasoning test of AA, K3 (Max) was 88.7 per cent and Pro was 86.3 per cent.
The official human walking demonstration is more visual: a virtual person starts to fall, undergoes several editions of training and is gradually able to walk. K3.1 Pro prepares simulators, arranges training and assessments, and then revises the methodology based on the results of the experiments. In the course of the process, the arm movement remained unsatisfactory and the adjustment incentive was not resolved; the model then examined the simulator, repaired 36 shoulder muscle arm computation errors and retrained the controller.
The movement of the human body is derived from simulation and control strategies, and Kimi is responsible for continuing to improve training based on results. It didn't stop at the adjustment incentive, and it found the error in the simulator, which was more interesting to me than the last one.
Such a long assignment would naturally increase the number of calls. Kimi brought down the API output price of regular Pro: token per million, K3 for $100, Pro for $68 and Highspeed for $136. Token is the unit in which the Platform calculates the amount of text and code usage, and the monthly users of the package are still using the package quota.
Kimi Code also consumes about twice as much quota as the high-speed version, and the speed increase only occurs at the model output stage. Time is still needed to read and write documents, execute orders and run programs. When normal Pro is fast enough, I'll leave quota for later changes.
quota may not be enough.
Some users think normal Pro is fast enough. He posted a call record amounting to 111 tokens per second, indicating that the high-speed version was adequate.
Chetaslua, who received an advance official visit from Kimi, posted a web-based presentation, describing itself as a single creation of the model, accompanied by the CodePen portal, where readers can see the effects directly.
There are also users who encounter quota problems when changing back and forth.
LanceYang gave the model five reference maps to make it an interactive 3D game handle. The first version was followed by two rounds, the shape of the component and the spatial relationship were still incorrect, and the model validation of the modification was subject to usage limits. He used a $199 package and had already used a portion of quota before starting the handle.
In a well-known pelican bicycle test, two users in another post reported very different consumption. A user with a $199 set of meals, a high-speed version and a top thinker, was created and asked to “continue to optimize” and roll quota was almost exhausted. Sanyeg uses a $599 package with the same model slot, but says that a pelican mission consumes only about 6 per cent. The mix, the task and the modification process are different, and it is difficult to judge which is more cost-effective only by proportion. So, as a whole, token consumption is a little fascinated.
According to Kimi ' s rules, monthly quota, models, context and tool call are all responsible for consumption, and Code also has roll-out limits. A user's ability to successfully complete a work depends on the full use from generation to modification.
There's a little bug, but it can fix itself.
I have chosen seven programming topics, each of which has a new session, allowing Kimi to run, check and modify himself, after which acceptance check will be delivered for the first time. We didn't change the code for it. Pictures, animations and prompt excerpts are attached to the complete prompt and input asset.
Pelican cycling has been measured too much, and there is fear of optimization. So we chose to let the bike ride the pelican.
Actual prompt (excerpt): The picture is shown on the other side of the same stage, on the other side, on the other side, on the other side, and on the right, on the other side, on the same side.... The bicycles are carried by pelican back, both wheels leave the ground, and pelicans walk in turn and forward.
Kimi made a left-right comparison: the left-left car, the right pelican walking, and the bicycle hanging over it. The position changed and there was nothing between the car and the bird's back, and it was not clear that pelicans were carrying bicycles. This part should be logically linked to the node in front of the bicycle, so it leads to a hiatus in the back, and another small problem is that the wheel of the bike is still spinning, which is somewhat problematic, but it can be optimized.
The physics is better. We set the ball to collide in a spinning seven-sided shape, observing actual collisions and operating patterns, and it worked really well.
Actual prompt (extract): a positive seven-sided radius of 320 in 0.2 radians/sec timewise. (a) 20 small balls with 12 radiuss, numbered 1 to 20, which do not overlap at the beginning and are gravitationally subject to vertical downward 500 pixels per second 2;
Through continuous operation of simulations, small ball positions were checked and no walling or ball throws were found. It also factored the speed of movement on the wall into the rebound. The first draft of the subject is already available to demonstrate gravity and collisions.
Sound analysis. I changed the bass and alto files to check if the spectrum really reflected the input.
Actual prompt (excerpt): Analysis of actual playable sound through Web Audio API, showing frequency spectrum, RMS noise and main frequency Hz; frequency axis range of at least 0 to 4,000 Hz, using interpretable real data and prohibiting pre-editing animation. Volume sliders affect output and analytical results.... with F05-demo.wav available for development, provided that the program supports other audio in the same format.
The chart corresponds to the frequency readings after the lower and upper pitch files. The volume is halved and the amplitude measured is nearly halved; the amplitude is zero and the main frequency value is no longer shown.
However, the short-duration sound will be on the map for a while before it disappears. It's enough to look at the sound, the strength and the strength, to keep the picture close to the drums and to adjust.
The pixel runners are only 12 x 8 pixels, and run in four frames.
Actual prompt (extract): Seven colour roads, stars and 12 x 8 pixels painted with up to 32 fixed RGB colours. The runner has a recognized four-frame running cycle, with the body always maintaining 12 x 8 logical dimensions; the road is rolling left and the role is at the centre of the image level.
Sizes, number of colours and action frames are met. The small man's arms and legs were twirled and the seven colours were rolled backwards, and although the picture was simple, the movement of running was already visible.
There is also a continuous deformation of the same card between buttons, players, etc.
Actual prompt (extract): The same card shell in the image centre becomes a button, loader, successful tick, player, volume control, switch, column chart, tip, then back to the button in 16 seconds.... a continuous transformation of shape, corner, size, icon and text, which cannot be replaced by a full page.
Most of the changes are smooth, and the problem is at the end: the text of the hint disappears, the picture leaves a blank button, and the text on the button suddenly appears at the beginning of the next round. Seeing a screenshot is easy to ignore and can be seen on a continuous basis.
In the field of measuring 3D modelling capabilities, we have chosen the case of the clear river map and have specifically asked people and ships to move. When crossing the bridge, the altitude below the foot will change along the bridge; the porter will first pick up the goods and then go to the dock and lay down. Bringing shot closer to see if it fits the scene.
Actual prompt (excerpt): Please create a 3D map of the Qingming River that can travel freely in the browser, with a focus on the river, the woody rainbow bridge, the bridgehead market, the shop along the river and the remote gate.... the person shall have identifiable physical and physical movements, continuing different activities, including at least walking along the street, crossing the bridge, pre-mortar trading, and moving cargo at the docks; the ship shall sail along the river and the person on board the ship shall move....walkers move along accessible roads, docks and bridges, with feet attached to the current surface or bridge heights, without entering rivers, buildings, stalls or crossing through bars.
During the production, I saw small people leaving the ground and people walking into the river and through the store. These problems all start out. And then Kimi optimised.
Kimi began his own gradual adjustment, the problem of modeling, the problem of people crossing the river, the basic logical questions that we did not add to prompt, and it modified the height of the person itself and changed the route through the shop.
Most of all, the shadows are very accurate, and it is clear that the models are beginning to understand the true patterns of the physical world.
These changes took place during Kimi ' s own inspection, and we did not add a hint.
In the delivery version, the moving union takes the goods to the dock and returns empty-handed and customers stop in front of the stall. In the near future, problems can still be found: when pedestrians move, the body enters the walls of a merchant. There is also a problem with the river lane: the surface is blocked by the ground, and the ship looks like a slide.
Finally, I had Kimi make a offline video clipper. Timelines, previews and export are subject to the same editing results and the new pages are maintained with the changes just made.
Actual prompt (excerpt): Each clip can set precise entry points, exit points; support splits, removes, rearranges, and undos and re-edits at the current top.... that you can really export a WebM video that can be played and allow real-time export. Export must reflect the cut and rearranged contents and cannot download original video impostors for export.... that saves the status of the item locally; after refreshing or reopening, you can restore the imported asset and edit results.
We'll take the runners and the pelicans, and we'll take each one, and we'll change it to the first, the second. Check the exported video, both in sequence and in interception range. New pages, asset and cut time lines have also been restored. It's normal to divide, delete, undo and redo in this round, and the first draft is ready for this simple clip.
There is also a minor problem with exporting documents: in the Edge browser used for testing, the video cannot be read for the first time, and it has to be broadcast once to be recognized. The video is correct and the file information is not yet complete.
Physical, audio and simple video clips are already useful in this cycle. Several visual features have also been handed over for further modification, and space relations and kinetic links still need to be checked.
Kimi referred to Pro as the first model of the K3.1 series and has not yet published Flash or Ultra. If the flash comes out later, I think it'll lower the cost of frequent calls.
Multiple Agents work at the same time, resulting in more model calls. If Flash can handle simple steps at lower prices and faster, developers can leave complex issues to Pro.
It was previously reported that recruitment of smart people had begun on the dark side of the month and that some robot company employees had received calls from Kimi. The report also refers to the statements of the interviewee that there are internal reservations about the Agenic Robotics (AI agent robots) route.
Google has tried to use Flash for robots. It was based on Gemini 3.5 Flash, who was trained in specific reasoning, and produced Gemini Robotics ER 2. The model is responsible for observing environmental, planning and monitoring tasks, which are then carried out by the lower action model.
The demonstration of human mobility, coupled with the recruitment of skilled personnel, was hard to believe, and Kimi wanted to continue exploring that area.
The best option I have at present is to involve strong models in the design and improvement of training so that control strategies can learn to act in simulations; when actually running, rapid models observe the environment, plan tasks, and then hand them over to lower control models. Training can be repeated and robots must respond in a timely manner when faced with changing circumstances. If Kimi had a low-delayed, low-cost multi-model Flash in the future, it could have been used to observe and plan this component.
As for Ultra, I expect it to add more difficult reasoning and long mission capabilities. Pro still has a gap in challenge testing without tools.
Although the test did not “stunned”, it could be inferred that Kimi had a relatively steady pace of advancement towards the real physical world and the overall optimization of capabilities, after all, on the K3 base.
Kimi has already handed over the papers, and now the open source is waiting for Qwen4.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.