China just dropped a new frontier model... and nobody can explain how
⚡ 速览
Fireship 实测 Moonshot 新发布的 3 万亿参数开源模型 Kimi K3.1：它在 Terminal-Bench 4.0 上距 Claude Opus 5.5 不到一分——如果 Anthropic 拦截 90–99% 蒸馏攻击的说法属实，这个成绩本不该存在。
视频顺带讲清了蒸馏悖论、被刻意删除的网络安全训练数据，以及那个唯一重要的基准：用 Three.js 生成 3D 马——Kimi 拿了 4.20 分，马最漂亮、动画最偷懒，仍落后于各家闭源前沿模型。
英文原文
Fireship tests Kimi K3.1 — Moonshot AI's new 3-trillion-parameter open-weight model that lands within one point of Claude Opus 5.5 on Terminal-Bench 4.0, a result that shouldn't be possible if Anthropic really blocks 90-99% of distillation attacks. The video walks through the distillation paradox, the deliberately removed cybersecurity training data, and the only benchmark that matters: building a 3D horse in Three.js, where Kimi scores a 4.20 — prettiest horse, laziest animation — behind the frontier closed models.
🗺 章节地图
💡 核心亮点
神秘模型 Kibitz 现身
名字头两个字母就出卖了它：LMArena 上的匿名前沿模型从一开始就是 Kimi。
上周，一个名叫 Kibitz 的匿名模型出现在 LMArena 并打出前沿水平成绩——Kibitz 是意第绪语，指下棋时站在你背后、支招让你把皇后送掉的那种人，Fireship 说这是编程 Agent 的完美比喻。名字头两个字母暴露了身份。
上周五 Moonshot 正式发布 K3.1：三万亿参数、开放权重，11 月 4 日开放完整下载——恰好在所有人以为 AI 泡沫破裂、跑分通胀到头的时候。
EN original
A mystery model called Kibitz
Last week an anonymous model named Kibitz — Yiddish for the guy behind you at a chess game giving tips on how to blunder your queen, which Fireship calls 'the perfect analogy for a coding agent' — appeared on LMArena performing at the frontier level. On Friday, Moonshot officially dropped Kimi K3.1, a new 3-trillion-parameter open-weight model that becomes fully downloadable on November 4, right when everyone assumed the AI bubble was popping and 'Trust Me Bro' benchmarks couldn't possibly get any better.
蒸馏悖论
如果 Anthropic 真拦住了 90–99% 的蒸馏攻击，一个这么强的中国开源模型本不该存在。
今年早些时候，Anthropic 指控 Moonshot、DeepSeek、MiniMax 等公司用成群假用户对 Claude 发动大规模蒸馏攻击，偷取输出与推理 token。
几周前 Anthropic 宣称来自中国的攻击已有 90–99% 被拦截；OpenAI 也展示技术连 PewDiePie 级别的小作坊蒸馏都拦得住。
大 AI 多年来暗示中国模型进步快靠的是偷输出——而自称"蒸馏人权论者"的 Fireship 反驳：大厂自己先蒸馏了人类集体大脑——我们的博客、GitHub、论文，以及被破坏性扫描的书（"搜 Project Panama 就知道"）。
EN original
The distillation paradox
Earlier this year Anthropic accused Moonshot, DeepSeek, MiniMax and others of running large-scale distillation attacks against Claude with swarms of fake users stealing output and reasoning tokens. A few weeks ago Anthropic claimed 90-99% of attacks from China are now blocked; OpenAI says its tech stops even small-scale distilleries like PewDiePie's. Big AI has long insinuated Chinese models improve fast because they steal outputs — but Fireship, a self-declared 'distillation is a human right' believer, notes the giants distilled the collective human brain first: our blogs, GitHubs, papers, and destructively scanned books ('just Google Project Panama').
把跑分读明白
信排行榜之前先问 harness：同一模型换个测试框架，分差可以差出天际。
K3.1 在 Terminal-Bench 4.0 上距 Opus 5.5 不到一分（Opus 以 66.4% 领跑），其他开源模型远远够不着。Artificial Analysis 独立复测拿到 67%，是他们测过的所有模型的最高分。
Kimi 还拿下 SWE Marathon——考验多小时软件工程任务的基准，相当于没有 AI 辅助时人类要干几周的活。但 harness 影响巨大：用不同框架跑，结果会差出天际。
EN original
Reading the benchmarks honestly
K3.1 is within one point of Opus 5.5 on Terminal-Bench 4.0 (Opus leads at 66.4%) and no other open model is remotely close. Artificial Analysis ran it independently and got 67% — the highest score of any model they tested. Kimi also wins SWE Marathon, a test of multi-hour software requirements — jobs that would take humans weeks without AI assistance. But the harness plays a big role: running models with different harnesses produces huge differences in results.
被故意切除的黑客能力
Moonshot 别无选择只能切除 K3.1 的黑客能力——这是美中 AI 博弈下的生存动作。
Moonshot 故意从训练数据里删掉了大量网络安全内容：内部未过滤的 checkpoint 能解出 90.5% 的"非法漏洞利用测试"，发布版只有 81%——明确目标就是让模型比 Pirate Software 还不会黑。
商业逻辑：几天前美国已禁止 DeepSeek 用于所有联邦民用机构，而监管方正愁没有全面封禁所有中国模型的理由——那将给 OpenAI 和 Anthropic 递上一场大胜。
EN original
Deliberately lobotomized hacking skills
Moonshot deliberately ripped cybersecurity data out of the training set. An unfiltered internal checkpoint solves 90.5% of their 'illegal exploit tests'; the release model only solves 81% — the stated goal was to make it even worse at hacking than Pirate Software. The business logic: days earlier the US banned China's DeepSeek from government use across all federal civilian agencies, and regulators 'would just love a good reason' to ban all Chinese models across the board — handing OpenAI and Anthropic a huge win.
Cyber Index 第一，靠对手弃赛
Kimi 的 Cyber Index 榜首含金量存疑：Claude 和 GPT-6 出于安全拒绝所有 prompt，直接弃赛。
即便被切除了黑客能力，K3.1 仍在 Artificial Analysis 新推出的 Cyber Index 上排所有公开模型第一。但要打星号：Claude 和 GPT-6 以安全为由拒绝了每一条 prompt——它们是弃赛而不是输了，所以这个排名更多反映拒答行为，而不是黑客水平。
EN original
Cyber Index #1, by forfeit
Despite the lobotomization, K3.1 still performed on top of all publicly available models on Artificial Analysis' new Cyber Index. The asterisk: Claude and GPT-6 declined every prompt on safety grounds — they forfeited rather than competed, so the ranking says more about refusal behavior than about hacking skill.
ThoroughBench：唯一重要的基准
Kimi 的马最好看，但动画靠扭曲 mesh 实现——好看不等于做得对。
Fireship 的私货基准：输入"用 Three.js 创建一匹逼真的可交互 3D 马"，再用"存在脑子里的专有损失函数"打 0–10 分。战绩：Gemini 4 Argon 0.69、GPT-6 Astra 5.21、Opus 6.39、Kimi K3.1 4.20。
Kimi 的马是样本里最漂亮的（配色精准；为跑这个 demo 他付了每月 250 元人民币的会员），但动画靠扭曲原始 mesh 实现——"敷衍了事"。Claude 的马建了 95,000 顶点的 mesh，35 根骨骼独立动画；Gemini Argon 也用了骨骼，只是不知为何多加了一根。
骨骼动画更难，但那是马走路不像被黑洞撕碎的原因。
EN original
ThoroughBench: the only benchmark that matters
Fireship's private eval: type 'create a realistic interactive 3D horse with Three.js', then score 0-10 with a 'proprietary loss function I keep in my brain'. Scores: Gemini 4 Argon 0.69, GPT-6 Astra 5.21, Opus 6.39, Kimi K3.1 4.20. Kimi produced the most beautiful horse of the sample (spot-on coloring; he paid 250 yuan/month to demo it), but its animation simply warps the original mesh — 'half-assed'. Claude built a 95,000-vertex mesh with 35 independently animated bones; Gemini Argon used bones too but added an extra one. Skeletal animation is harder — that's why the horse doesn't look torn apart by a black hole when it walks.
🎤 金句
Kibitz 是意第绪语，指下棋时站在你背后、给你支招让你把皇后送掉的那种人——说实话，这就是编程 Agent 的完美比喻。
EN original
Kibitz is the Yiddish word for that guy standing behind you at a chess game giving you tips on how to blunder your queen — which, to be honest, is the perfect analogy for a coding agent.
我认为蒸馏是一项基本人权，因为大 AI 公司蒸馏的其实是人类的集体大脑。
EN original
I believe distillation is a fundamental human right, because big AI companies have literally distilled a collective human brain.
在他们眼里，这是"大盗窃自动补全"。
EN original
To them, it's grand theft autocomplete.
如果我们信了 Anthropic 这份"我兄弟说的"蒸馏数据，那就意味着 Kimi 这些"我兄弟说的"跑分也不该存在。
EN original
If we can trust this Trust Me Bro distillation data from Anthropic, that would mean that these Trust Me Bro benchmarks from Kimi should not be possible.
这就是为什么 Moonshot 别无选择，只能给黑客能力做前额叶切除。
EN original
That's why Moonshot really has no choice but to lobotomize its hacking capabilities.
它的动画是靠扭曲原始 mesh 实现的——这是一种非常粗糙、敷衍了事的 3D 物体动画做法。
EN original
It achieves the animation by simply warping the original mesh, which is a very unsophisticated and half-assed way to animate a 3D object.
🧭 行动建议
- 1 付前沿价格之前先测 K3.1 02:15
一个距 Opus 5.5 不到一分（Terminal-Bench 4.0）的三万亿参数开源模型，11 月 4 日开放完整下载，且 Artificial Analysis 独立复测给出了他们测过的最高分。对预算敏感的编程 Agent 负载，它已经是正经候选——注意连 Fireship 跑 demo 都得买每月 250 元人民币的会员档。
EN original
A 3-trillion-parameter open-weight model within one point of Opus 5.5 on Terminal-Bench 4.0 becomes fully downloadable November 4, and Artificial Analysis independently measured the highest score of any model they tested. For budget-sensitive coding-agent workloads it is now a first-class candidate — note even Fireship needed a 250-yuan-per-month tier to demo it.
- 2 看到跑分先问 harness 02:31
不同框架跑出的跑分差异巨大，所以榜单数字描述的是"模型+框架"这个系统，而不是模型本身。做决策前，把厂商表格和 Artificial Analysis 这类独立复测分开对照。
EN original
Different harnesses produce huge differences in benchmark results, so a leaderboard number describes a model-plus-harness system, not a model. Compare vendor tables against independent runs like Artificial Analysis before making decisions.
- 3 安全相关场景别指望它 02:47
Kimi 的网络安全数据被刻意删减：发布版在内部漏洞利用测试上解出 81%，而未过滤 checkpoint 是 90.5%。做安全敏感评估时，发布的权重是被主动削弱的——把它算进你的威胁模型。
EN original
Kimi's cybersecurity data was deliberately stripped: the release model solves 81% of internal exploit tests versus 90.5% for the unfiltered checkpoint. For security-sensitive evaluations the published weights are intentionally weakened — treat that as part of your threat model.
- 4 地缘政治已是模型可用性风险 03:15
美国已禁止 DeepSeek 用于联邦民用机构，Moonshot 则抢先自削能力，不给监管者全面封禁的借口。如果关键工作流依赖中国开源模型，政策风险就该进架构决策的考量。
EN original
The US banned DeepSeek from federal civilian agencies, and Moonshot pre-emptively weakened its own model to avoid giving regulators a reason for a blanket ban. If a critical workflow depends on a Chinese open-weight model, policy risk belongs in the architecture decision.
- 5 验收实现方式，不只看成片 04:44
Kimi 的马最好看，但动画靠扭曲 mesh；Claude 的马没那么漂亮，却建了 95,000 顶点、35 根骨骼独立动画的正经骨架。评估 AI 生成的前端或 3D 产出时，记得按下播放键：构造是否正确体现在运动里，而不是第一张截图里。
EN original
Kimi's horse looked best but animated by warping the mesh; Claude's looked worse but shipped a 95,000-vertex mesh with 35 independently animated bones. When evaluating AI-generated frontend or 3D output, press play: correctness of construction shows up in motion, not in the first screenshot.
📖 全文
01 · 0:00 · Mystery model Kibitz and the K3.1 drop
0:00Last week, an anonymous mystery model named Kibitz showed up randomly on Chapot Arena performing at the frontier level. Kibitz is the Yiddish word for that guy standing behind you at a chess game, it giving you tips on how to blunder your queen, which to be honest is the perfect analogy for a coding agent. But it was pretty obvious who was behind this mystery model from the beginning because the first two letters of this word are a dead giveaway. Everybody knew it was Kimmy. Then right on Q last Friday, Moonshot dropped Kimmy K3.1, officially giving us a new three trillion parameter openweight model that will be fully downloadable on November 4th. Or so they say. And just when you thought the AI bubble was popping and the Trust Me Bro benchmarks could not possibly get any better, Kimmy made them better.
02 · 0:40 · The distillation war: Anthropic's blockade
0:40>> I can go lower. >> In today's video, we'll find out why Kimmy K3.1's performance should be impossible and put it to the test on the only benchmark that matters, the building 3D horses with 3JS. It is October 26, 2026, and you're watching the Code Report. Earlier this year, Anthropic accused Moonshot AI, the company behind Kimmy, alongside a bunch of other Chinese companies like Deep Seek and Miniax of running large-scale distillation attacks against Claude by sending swarms of fake users to steal Claude's output and reasoning tokens, which could then be used to train these open models that they give to us for free. And now, I believe distillation is a fundamental human right because big AI companies have literally distilled a collective human brain. but by taking our internet web blogs, our GitHubs, our scientific papers, and they even destructively scanned our books without asking. Just Google Project Panama. But OpenAI and Anthropic don't really see
1:29distillation as a human right like I do. To them, it's grand theft autocomplete. And they've invested millions and millions of dollars into stopping it. And just a few weeks ago, Anthropic said it pretty much solved the distillation problem with 90 to 99% of distillation attacks from China now being blocked. And over at OpenAI, they've proved that their tech can even stop small-scale distilleries like PewDiePie. But if that's true, that means an open Chinese model with Kimmy's performance should not be possible. Because for years, Big AI has insinuated that the reason Chinese models get so good so fast is that they steal Big AI's outputs. But if we can trust this Trust Me Bro distillation data from Anthropic, that would mean that these Trust Me Bro benchmarks from Kimmy should not be possible. like Kimmy K 3.1 is within one point of opus 5.5 on terminal bench 4.0.
03 · 2:15 · Benchmarks: Terminal-Bench 4.0 and SWE Marathon
2:15That's pretty crazy. And no other open model is remotely close. And we might even be able to actually trust these numbers because artificial analysis ran it independently and got 67% which is the highest score of any model they tested. But one thing to keep in mind is that the harness plays a big role here. Running different models with different harnesses will produce huge differences in benchmark results. In addition, Kimmy also wins on S. marathon, which is a test for multi-our software requirements, or in other words, jobs that would take humans weeks of work to finish without AI assistance. That's all
04 · 2:47 · The removed cybersecurity data
2:47pretty impressive, but there is one big caveat when it comes to coding and specifically cyber security. Moonshot deliberately ripped out a bunch of cyber security data from its training data set, where the goal was to make the model even worse at hacking than pirate software. >> I'm a genius, dude. And it worked because internally at Moonshot, an unfiltered checkpoint solved 90.5% of their illegal exploit tests. But the final release model only solves 81%. And that's a smart business move on their part because just a few days ago, the US banned China's Deepseek model from government use, including all federal civilian agencies. And they would just love a good reason to ban all Chinese models across the board to hand Open AAI and Anthropic a huge win and make all the insiders even more unbelievably rich than they already are. And that's why Moonshot really has no choice but to
3:32lobomtomize its hacking capabilities. But despite that, Kimmy still performed on top of all publicly available models on the new cyber index at artificial analysis. Although that's mostly because Claude and GBT6 declined every prompt on safety grounds and just took the L. But nowadays, the only AI capability that I
05 · 3:48 · ThoroughBench: the 3D horse test
3:48really care about is how well it can generate a 3D horse with 3JS. And to analyze the skill objectively and scientifically, I use a formal benchmark. It's a state-of-the-art eval I've developed after thousands of hours of research, which I call thoroughbench. The way it works is I literally just go into the chat window and type create a realistic interactive 3D horse with 3JS and then press enter. From there, I rank them from 0 to 10 based on a proprietary loss function I keep in my brain. Using this technique, Google Argon achieved a 0.69. GPT6 Astra achieved a 5.21 and Opus achieved a 6.39. But now it's time to try Kimmy K3.1. I tried to be cheap and use the free tier, but got denied. So now I'm paying 250 yuan per month to show you guys this demo. It was worth it though because Kimmy produced what is easily the most beautiful horse in this sample. Its coloring is spot-on. It looks realistic. And most importantly, its copulatory organ is a reasonable size. I was blown away by its beauty at first glance, but then I pushed play on the animation and found a problem. The animations are absolutely terrible. If we look closely, you can see it achieves the animation by simply warping the original mesh, which is a very unsophisticated and half-assed way to animate a 3D object. If we compare that
5:00to Claude, the horse itself may not look as beautiful, but it actually built a mesh with 95,000 vertices with each shape tied to one of 35 bones, and all those bones are animated independently. Even Gemini Argon used bones, but for some reason, it added an extra one. The skeletal animation is more difficult, but it's why the horse doesn't look like it's being torn apart by a black hole when it walks. So, the final score for Kimmy K 3.1 on Thoroughbench is a 4.20, which is pretty dang impressive, but
06 · 5:28 · Sponsor: Mobbin, and sign-off
5:28still behind the other Frontier models. But if you really want to build an amazing 3D horse or user interface, you need to check out Mobin, the sponsor of today's video. I've been using Mobin for over 5 years and its MCP server now lets your coding agent study its library of 600,000 real UI screens before it designs anything. So instead of asking for a login page and getting the same purple gradient vibe slop, you start by telling Claude Code, Cursor, or Codeex to study how the best apps built that complete user flow. It'll use Mobin to gather references and build a visual report of the patterns they all share. Then it'll recreate that flow and adapt it to your app. Every result points back to the original screen. This so you can check the source and after you connect Mobin, everything your agent designs will feel like it has actual taste. But try Mobin for free today and get a 20% discount at the link below. This has
6:18been the code report. Thanks for watching and I will see you in the next
01 · 0:00 · Mystery model Kibitz and the K3.1 drop
0:00上周，一个名叫 Kibitz 的匿名神秘模型随机出现在 LMArena 上，表现达到前沿水平。Kibitz 是意第绪语，指下棋时站在你背后、给你支招让你把皇后送掉的那种人——说实话，这简直是编程 Agent 的完美比喻。但它是谁家的，从一开始就很明显……这一点从名字的头两个字母就暴露无遗：人人都知道是 Kimi。然后掐着点儿，上周五 Moonshot 放出了 Kimi K3.1，正式带来一个全新的三万亿参数开放权重模型，11 月 4 日起将开放完整下载。至少他们是这么说的。就在你以为 AI 泡沫正在破裂、"我兄弟说的"式跑分不可能再惊喜的时候，Kimi 又把它们卷上了新高度。
02 · 0:40 · The distillation war: Anthropic's blockade
0:40「我还能更低。」在今天的视频里，我们要搞清楚为什么 Kimi K3.1 的成绩"本不该存在"，并用唯一重要的基准来实测它——用 Three.js 建 3D 马。现在是 2026 年 10 月 26 日，欢迎收看 Code Report。今年早些时候，Anthropic 指控 Moonshot AI——就是 Kimi 背后的公司——连同 DeepSeek、MiniMax 等一批中国公司——对 Claude 发动大规模蒸馏攻击：派成群的假用户去偷 Claude 的输出和推理 token，再拿去训练这些免费送给我们的开源模型。而现在，我认为蒸馏……是一项基本人权，因为大 AI 公司蒸馏的其实是人类的集体大脑——它们拿走了我们的网络博客、我们的 GitHub、我们的科学论文，甚至不经询问就把我们的书做了破坏性扫描，搜一下"Project Panama"就知道。但 OpenAI 和 Anthropic 可不这么看……
1:29在他们眼里，这不是什么人权，而是"大盗窃自动补全"。他们已经砸了数百万美元去阻止它。就在几周前，Anthropic 宣称蒸馏问题已基本解决：来自中国的蒸馏攻击如今有 90% 到 99% 被拦截。OpenAI 那边也证明了他们的技术连 PewDiePie 这种小作坊蒸馏都拦得住。但如果这是真的，那就意味着一个性能达到 Kimi 水平的中国开源模型本不该存在。因为多年来，大 AI 一直在暗示：中国模型进步这么快，是因为……它们偷了大 AI 的输出。可如果我们信了 Anthropic 这份"我兄弟说的"蒸馏数据，那就意味着 Kimi 这些"我兄弟说的"跑分也不该存在——比如 Kimi K3.1 在 Terminal-Bench 4.0 上距 Opus 5.5 不到一分。
03 · 2:15 · Benchmarks: Terminal-Bench 4.0 and SWE Marathon
2:15这相当疯狂，而且没有任何其他开源模型能望其项背。而且这些数字说不定还真可信——Artificial Analysis 独立复测拿到了 67%，是他们测过的所有模型中的最高分。但有一点要记住：测试框架（harness）在这里影响巨大。用不同的框架跑不同的模型，跑分结果会差出天际。此外，Kimi 还拿下了 SWE Marathon——一项考验多小时软件工程任务的测试，换句话说，就是没有 AI 辅助时人类要花几周才能完成的活儿。这些都……
04 · 2:47 · The removed cybersecurity data
2:47相当令人印象深刻，但有一个大坑，出在编程的特定方向上：网络安全。Moonshot 故意从训练数据集里删掉了一大堆网络安全数据，目标是让这个模型的黑客水平比 Pirate Software 还烂。「我是个天才，老铁。」而且还真管用：在 Moonshot 内部，一个未过滤的 checkpoint 能解出 90.5% 的"非法漏洞利用测试"，而最终发布版只能解出 81%。这是他们聪明的商业操作，因为就在几天前，美国……禁止中国的 DeepSeek 模型用于政府场景，覆盖所有联邦民用机构。他们正愁没有好理由全面封禁所有中国模型，好给 OpenAI 和 Anthropic 递上一场大胜，让那些内部人士变得比现在那份难以置信的富有还要更富。这就是为什么 Moonshot 别无选择，只能……
3:32给自己的黑客能力做前额叶切除。但即便如此，Kimi 在 Artificial Analysis 新推出的 Cyber Index 上仍排在所有公开模型之上——虽然这主要是因为 Claude 和 GPT-6 以安全为由拒绝了每一条 prompt，直接躺平认输。
05 · 3:48 · ThoroughBench: the 3D horse test
3:48但如今，我唯一真正在意的 AI 能力，是用 Three.js 生成 3D 马的水平。为了客观、科学地分析这项技能，我用一个正式基准——一个我经过数千小时研究开发出的最先进评测，我管它叫 ThoroughBench。它的玩法是：我打开聊天窗口，输入"用 Three.js 创建一匹逼真的可交互 3D 马"，然后按回车。接下来，我靠一个存在脑子里的专有损失函数给它们打 0 到 10 分。用这套方法，Google Argon 拿了 0.69 分，GPT-6 Astra 拿了 5.21 分，Opus 拿了 6.39 分。现在轮到 Kimi K3.1 了。我本想省钱用免费额度，结果被拒了。所以现在我每月花 250 块人民币，就为给你们做这个演示。不过值了，因为 Kimi 做出了这批样本里毫无争议最漂亮的马：配色精准、看起来很真实，最重要的是，它的交配器官尺寸合理。第一眼我就被它的美震住了，但接着我按下动画的播放键，发现了问题：动画一塌糊涂。仔细看会发现，它的动画是靠扭曲原始 mesh 实现的——这是一种非常粗糙、敷衍了事的 3D 物体动画做法。对比一下……
5:00Claude 的马：马本身可能没那么漂亮，但它真的建了一个 95,000 顶点的 mesh，每个形状绑定到 35 根骨骼之一，而且所有骨骼都独立做动画。Gemini Argon 也用了骨骼，但不知为什么多加了一根。骨骼动画更难，但正因如此，马走路时才不会像被黑洞撕碎一样。所以，Kimi K3.1 在 ThoroughBench 上的最终得分是 4.20——已经相当令人印象深刻，但……
06 · 5:28 · Sponsor: Mobbin, and sign-off
5:28仍落后于其他前沿模型。但如果你真想做出惊艳的 3D 马或用户界面，得看看 Mobbin——本期视频赞助商。我用 Mobbin 五年多了，现在它的 MCP server 能让你的编程 Agent 在动手设计之前，先研究它那 60 万张真实 UI 屏幕库。这样一来，与其要一个登录页然后收到千篇一律的紫色渐变 slop，不如先让 Claude Code、Cursor 或 Codex 去研究最优秀的 App 是如何构建完整用户流程的。它会用 Mobbin 收集参考，产出一份共性模式的可视化报告。然后复刻该流程并适配到你的 App，每个结果都回链原始屏幕，方便核对出处。接上 Mobbin 之后，你的 Agent 设计出来的东西会开始显得真正有品味。今天就去免费试用 Mobbin，走下方链接还能拿八折。本期……
6:18就是 Code Report。感谢观看，我们下期再见。
Fireship tests Kimi K3.1 — Moonshot AI's new 3-trillion-parameter open-weight model that lands within one point of Claude Opus 5.5 on Terminal-Bench 4.0, a result that shouldn't be possible if Anthropic really blocks 90-99% of distillation attacks.
The video walks through the distillation paradox, the deliberately removed cybersecurity training data, and the only benchmark that matters: building a 3D horse in Three.js, where Kimi scores a 4.20 — prettiest horse, laziest animation — behind the frontier closed models.
Fireship 实测 Moonshot 新发布的 3 万亿参数开源模型 Kimi K3.1：它在 Terminal-Bench 4.0 上距 Claude Opus 5.5 不到一分——如果 Anthropic 拦截 90–99% 蒸馏攻击的说法属实，这个成绩本不该存在。视频顺带讲清了蒸馏悖论、被刻意删除的网络安全训练数据，以及那个唯一重要的基准：用 Three.js 生成 3D 马——Kimi 拿了 4.20 分，马最漂亮、动画最偷懒，仍落后于各家闭源前沿模型。