信源说明：本文基于 Matthew Berman 频道视频《Deepseek crushed everyone》（视频 ID DDZ-a6bAn3Y，发布于 2026-10-28，全长约 21 分钟，转写共 11 段、每段 120 秒，覆盖 0–1271 秒）英文转写整理。同模型另见本博同日《五周、没升版本号、架构不变，DeepSeek V4.1-Flash-1017 却大幅变强》（Two Minute Papers 视角，含「自我反馈学习」推断与每任务成本提醒）；对比模型见《Kimi K3.1 完整评测》与《Kimi K3.1 上手实测》。文中全部测试结果、跑分与价格均为 Berman 自述或其引用的 Artificial Analysis 等面板数字，未独立核实；归属按「Berman 称／他引用／可以推断」分层。视频含 CoreWeave 广告段（240–360 秒段），本文不展开。

先说结论

Berman 的判断很直接（0–120 秒段）：DeepSeek 的新模型 V4.1 Flash「低调发布，但它是这个星球上最好的开源模型」，如果综合质量、价格与速度，「可能是当前整体第一的模型——不是质量最高，但接近」。支撑这个判断的不是跑分而是六组实测：3D 生成已经能造出可交互的机械表分解图与iPhone Duo 实体建模；多模态空间推理能数清 137 个 pastry 还会推算看不见的部分；代理浏览器能从一小时访谈里提取出 Berman 本人也会选的高光切片。真正的杀器在经济性：每任务成本 0.26 美元，是 Claude Fable 5.1 Max（7.63 美元）的约三十分之一。对读者的信号与 Two Minute Papers 那期互相印证：开源模型第一次在「性价比前沿」而不只是「追赶位」上，而比价的正确单位是每任务成本，不是每百万 token 单价。

六连测：3D、空间推理与代理能力

3D 机械表（0–120 秒段、120–240 秒段）：一句「做一个超写实的机械表 3D 图解，悬停显示零件名称与用途」得到单页应用「Mechanica」——秒针分针时针联动、光影随视角移动、点击游丝（hairspring）会放大并给出「淬硬钢制的扁平螺旋、不比头发丝粗」这类准确解释，还能整体爆炸分层展示。唯一缺陷是表背上没有自动陀（上弦摆陀），Berman 认为这是全程最接近「瑕疵」的地方。Fall Guys 克隆（120–240 秒段）：HTML 单页的「Bean Blitz」可玩——锤子能把角色砸下台、空格扑倒、有碰撞穿模但整体成立；他的评级是「不如 Opus 5.5，但相当接近，而价格只是零头、速度极快」。iPhone Duo 建模（240–360 秒段）：模型自己上网查规格再 1:1 重建——Apple logo、屏幕反光、内屏磨砂质感、按键位置、铰链全部正确，折叠动画有出屏小毛病；同一测试他给新 Kimi 模型做过，结论是 DeepSeek 明显更好。

多模态计数（360–600 秒段）是最惊艳的一段。他上传一张摆满 pastry 的柜台照片问「有多少个」：模型花 4 分 25 秒逐个清点，报出 137 个并全部命名（橙子马卡龙 24 个——Berman 自己数是 25 个，他认为「有得争」；黄油可颂 8、奶油角可颂 9）。关键是它主动声明：马卡龙、可颂、布朗尼是堆叠的，一部分被遮挡，「整柜的可靠区间是 130–145」。追问「把看不见的也估进来」后，它用几何推理把马卡龙金字塔从可见 24 个修正为总数约 45 个（锥形堆、底宽 5–6 排、5–6 层，可见三角面 24 意味着底层 19–25），可颂角堆按 2–3 层深给可见数加约 60%——追加计算只用了 23 秒。Berman 称这是「恒星级的视觉理解」（全程开最大推理档）。代理浏览器（600–840 秒段）：让它审看他的 Sam Altman 访谈、提取五个「最值得切片」的时刻并给精确时间戳，6 分 24 秒完成；开源权重的好处是全部思维链与工具调用可见。第一选「我一年没看过一行代码了」（时间戳经他核验无误，含「革命战争重演」那个金句），第三选「意识辩论是夺权的后门」——Berman 称这五个全是他自己会选的。

本地跑通：开源权重与「一人软件」时代

性能之外，Berman 更兴奋的是拥有感。他在自己的 DGX Station 上完全离线跑这个模型（720–840 秒段），并强调量化版「一块还不错的高端消费级显卡」就能跑。两个本地产物：其一是 50 分钟造出的离线优先 Excel 克隆 Gridline——公式（=B7+C7）、拖拽填充、图表、拖放全部可用；其二是个人 PDF 阅读器 Folio（840–1080 秒段），动机很真实：他不想再用「臃肿、塞广告」的 Adobe Acrobat。这里出现一个值得抄的提示词接力工作流：他先让 Opus 5.5 把需求写成一份极详细的规格（Electron+React+TypeScript、PDF.js 渲染、暖白/近黑双主题、批注/签名/相机翻拍签名、任意 PDF 一键暗色模式），再把这份提示词原样粘给 DeepSeek 执行——前沿模型出规格、廉价模型出代码，本地跑基本零成本，走 API「也就一两美元，换来一个用一辈子的 PDF 阅读器」。他的口号是：你现在就该给自己写软件——要用的工具不存在或很难用，不再是停下的理由。

性价比账本：Pareto 前沿与每任务成本

跑分部分（1080–1271 秒段）他给三组数字，全部引自其展示面板（转写称 Artificial Analysis 等）：DeepSweep v1.1 的质量-成本 Pareto 前沿上，DeepSeek 位于最优位置，压过 Opus 5.5、Fable 5、GPT-6.1-Sol、Astra-6-Sol、Gemini 3.8 Flash、GLM-5.3-Max 等一众旗舰；Humanity’s Last Exam 42 分，与 Kimi K3、GLM-5.3 同档；Terminal Bench 42，Automation Bench、Agents Last Exam 也都靠前——但他提醒「跑分只讲了一半故事」。第三组是他认为看新模型最重要的指标：每智能任务成本（单价 × 完成任务所需 token 数）。阶梯是：Claude Fable 5.1 Max 7.63 美元居顶，Qwen 3.8 Max 5 美元，GPT-6-Astra 3.26 美元，Kimi K3.1 Pro 2.11 美元，而 DeepSeek 是 0.26 美元——「便宜得离谱，而且模型真的很好」。托管侧：转写称 Lithos AI 报价每百万输出 0.60 美元、输入 0.15 美元、缓存命中免费，推理速度最高 1,286 tokens/秒；Base10 约 845 tokens/秒、输出 1.20 美元、输入 0.30 美元、缓存命中每百万仅 0.01 美元——他认为 Base10 是当前综合最优的供应渠道。这与 Two Minute Papers 主持人同一周的提醒严丝合缝：那期指出 DeepSeek 同任务可能多耗两到三倍 token，「每任务成本」才是可比单位——两位主持人在没通气的情况下收敛到了同一个结论。

榜单与体感的裂缝

结尾藏着本期最有张力的承认（1200–1271 秒段）：Kimi K3.1（Moonshot 最新模型）在榜单总分上更高，但 Berman 个人判断 DeepSeek「好得多」，并建议观众自己去对比。这与他 10 月底的 Kimi K3.1 评测（彼时该模型以跑分惊艳、实测有翻车）构成一组镜像：榜单赢的不一定上手赢。对依赖公开榜单做选型的人，这条裂缝比任何单项跑分都值得记录——尤其是当两个模型价差在每任务 0.26 对 2.11 美元这个量级时，「自己去跑十个真实任务」的成本已经低到没有借口。

归属与转写说明

本文全部事实与数字出自本集转写（含 137 个 pastry、每任务 0.26 美元、845/1286 tokens/秒等 Berman 自述与其引用的面板口径，未独立核实）；「最好的开源模型」「可能是整体第一」均为 Berman 个人判断，非独立测试结论。转写校正：DeepSeek V 4.1→V4.1、Opus Five→Opus 5.5、Fable Five Point One Max→Fable 5.1 Max、GPT Six Point One, Soul→GPT-6.1-Sol、Astra Six, Soul→Astra-6-Sol、Gemini Three Point Eight, Flash→Gemini 3.8 Flash、GLM Five Point Three Max→GLM-5.3-Max、Quen Three Point Eight Max→Qwen 3.8 Max、Kimi K three point one→Kimi K3.1、Base 10/Base Ten→Base10、cash hit→cache hit；按原文保留：Lithos AI、Mechanica、Bean Blitz、Gridline、Folio、Arya（CoreWeave 助手）、DeepSweep v1.1。附录转写已按确认项机械替换。

转写原文（修正版）

[0s - 120s] DeepSeek’s latest model flew under the radar, and it is the best open source model on the planet. The first thing I wanted to test combines a few different capabilities, including its ability to generate three D models, spatial awareness, internal knowledge the model has in its weights. And I think this is such a cool test. So I said, “Build a three D mechanical watch diagram so we can see exactly how a mechanical watch works. Make it hyper realistic and beautiful. When the user hovers over a piece, add information about what it’s called and what it is used for. And here’s what it came up with. This is the Mechanica, and look how beautiful it is. And if you look closely, all of the pieces work. They fit correctly. These are tiny little pieces, and it looks three D. You can see the lighting. You can see the shadows moving as I’m moving the watch face. You can see the seconds hand, the minutes hand, the hours hand. Everything looks really good. Now, if I hover over any of these things, it tells me what it is. I can click over the hairspring. So if I just hover there, it says hairspring, a blued spiral spring that returns the balance and sets the watch’s rhythm. If I click into it, it kind of highlights and centers it. So the hairspring is a flat spiral of hardened steel, no thicker than a human hair. It stores the balance’s kinetic energy at the end of each swing and gives it back on the return. So I can say, “Examine closely.” It zooms into it. Look how awesome this is. It gives me some data about it. I can actually click into each of the systems right here, and then the overall parts of the watch right here. Go back to the overview, and we can spin it around. We can look at the back. The one thing I don’t see here is the little weight on the back where you move your wrist and it moves around and charges the watch. I actually am forgetting in this moment what that’s called, but nothing is here in the back, so that was a little disappointing. But otherwise, I think this is incredible. Now, check this out. If I

[120s - 240s] Click explode, it actually explodes it, and you can see each layer of the watch completely exploded out, and you can still hover over it. And of course, all of this was built in a single web page, and I just think this is absolutely beautiful. The detail, the spatial awareness, its three D generation, the lighting, the shadows, everything looks really good, with the exception of the back, which is the only improvement I would have liked to see. All right, the next thing I had it create is a very popular multiplayer game called Fall Guys, and this is one that I’ve done a number of times before. So while I show you this, I’m going to show you the other versions that other models have created. And so the prompt for this was quite simple: build a Fall Guys clone. All right, so this is called Bean Blitz, and let’s play. So immediately I see it looks really good. The three D generation is really good. This is again just built with HTML. Oh, I got knocked off. Okay, so yeah, everything three D looks really good. You can see some clipping of the different objects. This doesn’t actually do anything, so it looks like it would be spinning, but it’s not. So that’s okay. I can’t fall in the acid right here. So let’s keep jumping. Now the hammers look good. Let’s see if it actually knocks me off. It does. Boink. I can dive hitting the spacebar. Okay, fell off. Now we have these balls, and I lost. So I think overall, I’m really impressed with its ability to clone this game. It isn’t as good as Opus 5.5, but it’s quite close. And keep in mind, this model is a fraction of the price and very very fast. All right, next I wanted it to create a three D model of the iPhone Duo, which requires it to go out on the web, understand the spec, understand the design, and actually be able to recreate it in three.

[240s - 360s] Space and I think it absolutely crushed this test. So here it is, and I think this is flawless. The Apple logo looks perfect. Look at that; you can even see the reflection of the light in the screen. This side looks matted. It is a matte finish on the inside screen. All the buttons are in the right place. The hinge looks correct. Everything is flawless. And by the way, I did this exact test with the brand new Kimi model. Let me put that on screen because it shows DeepSeek is much better. So then you can actually fold it. Now there’s a little issue with the folding animation because when you click it, it kind of goes off screen. So that is an issue with this simulation. But let’s fold it. There we go. And so the actual screensaver disappears for a second while you fold it. But overall, I think this looks incredible. And if you’re thinking about using this DeepSeek model in production for an app you’re building, you should definitely make sure it is production ready using the sponsor of today’s video, CoreWeave. CoreWeave Arena lets you load test your AI before you ship to production. Here’s what that looks like in Forge, CoreWeave’s AI development platform. It starts with one request at a time, then five, then ten. I can see the response time, the throughput, and estimated cost. When ten requests came in, it took longer than five. But why? That’s when I go to Arya, the AI assistant in Forge, to help me figure it out. It recommends reducing unnecessary processing for each answer. Now I’m going to test that before I launch. Arena is a guided program for testing Forge infrastructure or both with CoreWeave’s engineers. CoreWeave is the essential cloud for AI. CoreWeave Arena gives you the insights you need to know on what will work, what will break, and what it takes to get your AI into production. So apply for Arena Pass today.

[360s - 480s] I’m going to drop a link down below. Let them know I sent you and check it out. And this new DeepSeek model V4.1 Flash is also multimodal, which means it does accept images. And so I uploaded this image, and it is a bunch of different pastries. And I wanted to see how well it understands what’s in the image. And this is a really hard task. So check this out. How many pastries are on the counter? Count them individually. I went through the counter front to back. It took four minutes and 25 seconds in total, and counted every distinct item I could. Here’s my tally: 137 individual pastries visible on the counter. Number one, orange macarons, and it is in the pyramid, and it counted 24. So if we look at it, this is what it’s seeing right here: one, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen, eighteen, nineteen, twenty-one, twenty-two, twenty-three, twenty-four, twenty-five. Now I count 25 visible, but 24 is very close, and it’s arguable. Now there are a bunch of macarons that you cannot see in this image, and wouldn’t it be nice if this model could understand what is not being shown in the image just as well as what is being shown? I’m going to get back to that in a second. Here we go: butter croissants, eight; cream-filled horn croissants, nine. So not only was it able to count every single pastry in the image, but it was actually able to name them all, which I find very impressive. Now, without me asking it, it said the macarons, croissants, and brownies are stacked, so a handful are partly or fully hidden. Those three groups are the fuzzy ones. A defensible range for the whole counter is roughly 130 to 145. So it immediately recognized that there are likely pastries behind the ones that are stacked that cannot be seen. So I asked it now: estimate how many are on the table in total, include the ones that you might not be able to see directly. For example, there are probably more macarons.

[480s - 600s] Behind the ones that are visible directly in the pyramid. So here we go. Now it has the list of visible and hidden. So it estimated another 21 hidden of those macarons. Here’s butter croissants on the back stand. Let’s take a look. So here we go. One, two, three, four, five, six, seven, maybe seven and a half visible, and it’s saying there are five more that you cannot see that are stacked behind it. Look at its spatial awareness here. Macarons are the big ones. It reads as a cone-shaped heap, about five to six wide at the base and five to six layers tall. A triangular front face of 24 implies a base layer of roughly 19 to 25 tapering up. So 45 total is much more likely than 24. Here we go. The croissant horn piles, two to three layers deep. So I added about 60% to the visible count. I am beyond impressed with this. And the cool thing is, the original count only took four and a half minutes, and the follow-up count, it had already done the calculation. It only took another 23 seconds. And by the way, this is on max reasoning. So stellar visual understanding with this DeepSeek model. All right, the next thing I want to show you is how fast this model is. It has the moniker Flash, which means it’s going to be very fast, but it is surprisingly fast. So what we’re going to do, we’re going to select DeepSeek V4.1, and I’m going to turn off reasoning effort, and I’m going to say, write a 1,000 word story in this chat window. Do not use tools. Now the reason that I do that is because even if you have the thinking effort set to off, and you tell it to write a thousand word story, it’s still going to think, okay, maybe I should spin up a text document. Maybe I should do something else. All I wanted to do is output the text directly in this window. So let’s send it and let’s see how fast it goes.

[600s - 720s] Look at that. That’s probably 125, 150 tokens per second, and that’s not the fastest you can get it. In fact, I’m going to show you this in a little bit. This is artificial analysis. You can get this model up to 1,286 tokens per second from Lithos AI. Base10 is serving it at 845 tokens per second, so you can get very fast inference with this model. That is why I’m thinking this might be the current best overall model on the planet. Not the highest quality, but close. But with its cost, its quality, and speed, I think it’s number one. Okay, next, I wanted to test its ability to a use a browser, b understand how to specifically use YouTube, c look at a video and understand what a clipable moment in an interview is, and then d give me the correct timestamps. So I gave it my Sam Altman interview. I said, review the YouTube video and extract the top five most clipable moments. Make sure to give me exact timestamps and why you think it is worth clipping. Then I just linked to the video. So it worked for six minutes and 24 seconds. And you know, it’s really cool with an open source, open weights model. You can see all of the chain of thoughts, all of the tool calls. So the first clip it decided on, I haven’t looked at a line of code in a year, and it gave me the specific timestamp, which is correct. I verified it. It gave me a link. It told me the exact quote. I realize I have not read or even looked at a line of code in probably a year. I don’t think I’ll ever do that again. Or if I do, it’ll be like a Revolutionary War reenactment when I’m old and bored and trying to remember the glory days. And that’s Sam Altman’s quote. Why did it choose that? The CEO of OpenAI admitting he hasn’t read code in a year is inherently newsworthy. And the Revolutionary War reenactment line is a vivid, funny image that travels well outside the AI bubble. Self-deprecating.

[720s - 840s] Zero setup required, perfect shorts, TikTok length, hook text. Sam Altman hasn’t read a line of code in a year, so that is a very good choice in my opinion. Next, coders have never been busier. Another great choice. Three, the consciousness debate is a backdoor to taking power from people. Another really good clip, one of my favorite from that interview. I get very nervous when people start using the AI consciousness debate to push into. Well, maybe people shouldn’t always be in control of the future. The future has to be about people. And so, yeah, these are all clips that I would have personally chosen. And so, the next thing I would probably give it, although I haven’t done it yet, is to actually download the video and clip it itself. And I think it can do that, but that’ll have to be for another video. All right, here’s the next one: a full Microsoft Excel clone. And I want to say something about this before I share it with you. I loaded the DeepSeek model locally. Now, granted, I have a DGX station, which most people don’t have. It’s very expensive, but you can get a quantized version of DeepSeek on a decently high-end consumer-grade GPU, and it makes me so excited. So I ran this and built this completely locally. So all I said to do was build an Excel clone for the browser. It worked for 50 minutes, and it built Gridline, an offline-first Excel clone. So if I click into it, here it is. It doesn’t have the full functionality of Excel, but it is very good nonetheless. So if I double-click into here, 23 plus 44 equals 67. I can say 44 here equals B7 plus C7, and it does it correctly. It has bold. I can drag this down. After dragging it down, it basically continued that formula into each of the cells that I dragged it to. So here I insert a chart, and it has a.

[840s - 960s] Beautiful chart, very simple. I can visualize another chart like so. There’s another one. Drag and drop, very easy. And next, I gave it the Rubik’s Cube test. Build a 3D simulation of a Rubik’s Cube scramble and solve. Very simple prompt, and there’s a reason I gave it a very simple prompt. Is it able to build the thing that I have in my mind? The thing that we all know what it should look like, but without me having to explicitly tell it every single nuance of the task. And here it is. This Rubik’s Cube so far looks really good. All the sides look good. There’s white under here, blue, orange, yellow, green, red. Scramble. All the colors seem to be sticking to the appropriate sides. Everything still looks good after scrambling. We can actually see the moves scrambled right here. And then if I click solve, everything looks good. The rotations look accurate. The physics of the cube looks accurate. And yes, it solved it. So so impressive. This is an open source, open weights model that you can download and do whatever you want with. And it is really good. Next, you should be building your own software at this point. I have built multiple apps recently that I just needed in my life, and it either didn’t exist or the app that existed was garbage, and I wanted to create something that was custom to me. One example is a PDF reader for Windows. No, I do not want to use Adobe Acrobat. It is bloat. It is spammy, and I do not need 90% of the features it has. So I built my own, and you can too. I had originally built this with Opus 5.5, and I asked Opus 5.5 to give me a detailed prompt, very different from what I’m used to doing with these tests, but give me a detailed prompt, and then I’ll copy paste the prompt into DeepSeek. So I have it, DeepSeek, high thinking settings, build a complete PDF reader and annotation app called.

[960s - 1080s] Folio from scratch. It’s going to be using Electron, React, and TypeScript. PDF.js for the PDF rendering. It told it exactly how I want it to look and feel: warm, off-white light theme and near-black dark theme. Subtle borders, compact controls. Here, it told it exactly how I want the PDF reading to be: drag and drop, you can highlight text, etc. And so, after all that, it is a very detailed prompt. And here it is. Here’s Folio. This is a beautiful, simple PDF reader and editor, and it just works. So, some of the features: you have the pages, the table of contents here. You can click around. You can hide that window if you want. You can zoom. All very like basic stuff so far. You can draw on it. You can have different colors. You can move the drawing around. You can actually change the color just like so. You can add a text note. You have signatures, so you can actually draw a signature. You can upload a signature. You can take a picture, so you can like actually draw a signature on a piece of paper, hold it up to the camera, it will take a picture, flip it, reverse it, and then you can use that signature anywhere you want. It also has dark mode, which I appreciate. You can just make any PDF into dark mode, so it’s easier on your eyes. Saving, you can export, fit within the page, you can rotate it clockwise, you can print it, everything you would expect from a PDF reader, and it was basically free to make. I mean, I ran it completely locally. If you’re paying an inference provider for this, you’re probably paying less than a couple dollars to make a PDF reader that you can have for the rest of your life. It is so impressive. So, from these demos, hopefully, you can see why I’m calling this potentially the best overall model on the planet right now. There hasn’t been another open source model that I have been this impressed with, and it’s not just that it’s fast and cheap, and you can host it locally. You can

[1080s - 1200s] Get inference at over a thousand tokens per second. It’s the quality. It is actually high quality, actually high intelligence. And in fact, let me show you some of those benchmarks and stats. Now, here is one of the most important benchmarks you need to look at. This is DeepSweep v1.1, and what you’re seeing is this is the Pareto frontier on the quality cost combined metric. It is the best model, and so you can see that right here. Here’s a bunch of other models. Here’s Opus 5.5, Fable 5, GPT-6.1-Sol, Astra-6-Sol, Gemini 3.8 Flash, GLM-5.3-Max, and up here in the top left, this is where you want to be. Higher up means better. To the left means cheaper, and it is beating every single other model out there. I mean, this is super impressive. This model is the best bang for your buck, and you’re not really sacrificing all that much quality unless you’re doing the most complex tasks ever. Here are some other benchmarks we have: Humanity’s Last Exam. It’s coming in at a 42, which is right on par with Kimi K3 and GLM Five Point Three. Here’s Terminal Bench coming at a 42. Automation Bench, Agents Last Exam. It just scores really well. But of course, benchmarks only tell part of the story. And now that you’ve seen it actually in action, you should be using this model. And I want to show you one of the most important metrics that you need to be looking at anytime you see a new model, which is cost per intelligence task. And that combines the cost of the model with how many tokens it takes to complete a given task. And so you want the lowest cost, and you also want the highest density of intelligence, which results in less tokens being used per task. And so at the very top, maybe unsurprisingly, Claude Fable 5.1 Max at seven dollars and sixty three cents. I’ll show you a couple others. Qwen 3.8 Max at five dollars, very expensive. Here’s Astra

[1200s - 1271s] At three twenty-six, here’s Kimi K3.1 Pro at two dollars and eleven cents. This model just came out, and I think DeepSeek is better than it. And all the way down here, look at this: twenty-six cents per task completed. That is so inexpensive, and it’s a very good model. And speaking of price, here’s what you can get the model at on Lithos AI. You can get it for sixty cents per million output tokens, fifteen cents per million input, and it’s free if it’s a cache hit. Free. Here’s Base10, probably with the overall best offering right now because remember it gets about eight hundred tokens per second, and it’s a dollar twenty per million output, thirty cents per million input, one penny per million cache hit tokens, one penny. I’m so impressed with DeepSeek. Keep shipping the goodness, DeepSeek. Thank you. And I mentioned Kimi K3.1 a few times, and it’s the newest model from Moonshot, and it scored higher overall in the benchmarks. But I personally think DeepSeek is much better. You should make the comparison yourself. Go watch my Kimi video here.