WEBVTT 00:00:00.000 --> 00:00:04.660 DeepSeq v3 .2, an open source model, just scored 00:00:04.660 --> 00:00:07.480 a gold medal level on the International Mathematical 00:00:07.480 --> 00:00:10.800 Olympiad. That one fact, that single data point, 00:00:12.080 --> 00:00:14.140 it basically changes everything we thought we 00:00:14.140 --> 00:00:16.739 knew about the AI arms race. It's really hard 00:00:16.739 --> 00:00:19.600 to overstate what a big deal that is. For years, 00:00:19.699 --> 00:00:21.539 we've all just kind of assumed that only the 00:00:21.539 --> 00:00:24.600 trillion dollar companies, big players, could 00:00:24.600 --> 00:00:27.379 build this kind of state of the art AI. The source 00:00:27.379 --> 00:00:29.480 material we're looking at today proves that assumption 00:00:29.480 --> 00:00:31.780 is now, well, it's history. Open source isn't 00:00:31.780 --> 00:00:33.780 just catching up anymore. In some key areas, 00:00:33.780 --> 00:00:36.000 it's actually leading. It's leading. Exactly. 00:00:36.539 --> 00:00:38.939 So our mission in this deep dive is to get you 00:00:38.939 --> 00:00:41.159 straight to what matters in these sources. We're 00:00:41.159 --> 00:00:42.799 going to unpack the two different versions of 00:00:42.799 --> 00:00:46.039 V3 .2. We'll review those frankly shocking benchmark 00:00:46.039 --> 00:00:49.000 scores against models like GPT -5 .2 and Gemini. 00:00:49.320 --> 00:00:51.039 And then we'll get into the technical magic. 00:00:51.460 --> 00:00:54.210 Yeah, that deep -seek sparse attention. That's 00:00:54.210 --> 00:00:56.590 a big piece of the puzzle. And we'll finish up 00:00:56.590 --> 00:00:59.270 with the real world coding test that prove it's 00:00:59.270 --> 00:01:02.490 not just scary, and maybe most importantly, the 00:01:02.490 --> 00:01:05.170 price tag that is already sending waves through 00:01:05.170 --> 00:01:07.129 the whole industry. It's a lot to get through, 00:01:07.150 --> 00:01:09.510 so let's just dive right in. Okay, so let's start 00:01:09.510 --> 00:01:11.549 with the strategy. Because the first thing that 00:01:11.549 --> 00:01:13.760 jumps out... from the sources is that they didn't 00:01:13.760 --> 00:01:17.560 just release one model, they launched V3 .2 in 00:01:17.560 --> 00:01:20.920 two distinct flavors. Which was such a smart 00:01:20.920 --> 00:01:23.519 move, so user -focused. They really tailored 00:01:23.519 --> 00:01:25.299 the models for different needs, which is something 00:01:25.299 --> 00:01:27.689 you don't always see done this well. First, you 00:01:27.689 --> 00:01:30.489 have the standard V3 .2. You can think of this 00:01:30.489 --> 00:01:32.510 as, I don't know, the reliable everyday car. 00:01:32.650 --> 00:01:34.750 It's efficient, it's fast, and it handles common 00:01:34.750 --> 00:01:37.510 tasks perfectly. You know, drafting emails, summarizing 00:01:37.510 --> 00:01:40.310 articles, basic coding. It's built to be cheap 00:01:40.310 --> 00:01:42.969 and effective. And then you have the Jeep Seek 00:01:42.969 --> 00:01:47.120 V3 .2 Speciale, this one. This is the race car. 00:01:47.319 --> 00:01:50.239 It's built for heavy complex thinking. If you're 00:01:50.239 --> 00:01:52.560 tackling a really tough math proof or a complex 00:01:52.560 --> 00:01:54.859 engineering problem that needs deep multi -step 00:01:54.859 --> 00:01:57.840 reasoning, you bring out Speciale. And the sources 00:01:57.840 --> 00:02:00.060 are pretty clear on how it does that. The Speciale 00:02:00.060 --> 00:02:03.540 model just, it allocates more compute, more power 00:02:03.540 --> 00:02:06.599 to think deeper before it even writes a single 00:02:06.599 --> 00:02:08.900 word. But what's interesting is that they both 00:02:08.900 --> 00:02:11.520 share the same core architecture. The sources 00:02:11.520 --> 00:02:14.139 call them reasoning first models. So unlike a 00:02:14.139 --> 00:02:16.699 lot of older models that are just sort of guessing 00:02:16.699 --> 00:02:19.199 the next most likely word. Right. And that's 00:02:19.199 --> 00:02:21.639 when they can start to hallucinate or just break 00:02:21.639 --> 00:02:26.080 down logic. Precisely. This model, V3 .2. It 00:02:26.080 --> 00:02:28.479 actively tries to understand the logic and the 00:02:28.479 --> 00:02:30.599 structure of your question before it generates 00:02:30.599 --> 00:02:32.719 an answer. It builds a logical framework first. 00:02:33.199 --> 00:02:35.280 So for a developer listening to this, how big 00:02:35.280 --> 00:02:38.139 of a deal is this dual model approach for balancing 00:02:38.139 --> 00:02:41.280 API costs versus raw power? It lets you optimize 00:02:41.280 --> 00:02:43.539 performance directly. You only pay for that race 00:02:43.539 --> 00:02:45.520 car processing when your task actually needs 00:02:45.520 --> 00:02:47.560 it. OK, this is where the story gets really, 00:02:47.560 --> 00:02:49.259 really good. We have to talk about the benchmarks 00:02:49.259 --> 00:02:51.979 and the sources zeroed in on the ultimate test. 00:02:52.219 --> 00:02:55.680 the IMO, the International Mathematical Olympiad. 00:02:55.979 --> 00:02:58.300 And the IMO. It's not just some hard high school 00:02:58.300 --> 00:03:02.419 exam. It's designed to require creative, novel 00:03:02.419 --> 00:03:04.680 problem solving. You can't just memorize formulas. 00:03:04.879 --> 00:03:07.620 Most AI just fail spectacularly at it because 00:03:07.620 --> 00:03:09.219 they don't have that deep reasoning. And yet 00:03:09.219 --> 00:03:12.020 the special version. It achieved a gold medal 00:03:12.020 --> 00:03:14.080 level score, not just passing, but performing 00:03:14.080 --> 00:03:17.219 at the absolute elite level of the smartest high 00:03:17.219 --> 00:03:19.469 school students on the planet. I mean, that is 00:03:19.469 --> 00:03:21.770 a massive validation of their whole design. And 00:03:21.770 --> 00:03:23.009 you really have to look at the head -to -head 00:03:23.009 --> 00:03:24.870 numbers from the source material to get it. They 00:03:24.870 --> 00:03:26.949 put it up against the best closed source models 00:03:26.949 --> 00:03:30.250 out there. Let's start with the AME 2025 math 00:03:30.250 --> 00:03:33.330 test. Special didn't just compete with GPT 5 00:03:33.330 --> 00:03:36.969 .2 high. It beat it. Special scored a 96 points 00:03:36.969 --> 00:03:41.729 RO. GPT 5 .2 got a 94 .6. That's a clear statistical 00:03:41.729 --> 00:03:44.370 win in pure logical reasoning. And it wasn't 00:03:44.370 --> 00:03:47.030 a fluke. On graduate level science, on the GPQA 00:03:47.030 --> 00:03:49.870 diamond test, it topped. with GPT 5 .2. Then 00:03:49.870 --> 00:03:52.409 you look at coding on live codebench and 88 .7, 00:03:52.449 --> 00:03:54.129 that puts it right up there shoulder to shoulder 00:03:54.129 --> 00:03:57.189 with Gemini 3 .0 Pro. So if they're outperforming 00:03:57.189 --> 00:03:59.150 a top -tier model in advanced math, what does 00:03:59.150 --> 00:04:01.189 that really tell us about the quality of Deep 00:04:01.189 --> 00:04:03.389 Six Core reasoning architecture? It tells us 00:04:03.389 --> 00:04:05.569 the reasoning first design philosophy works. 00:04:06.229 --> 00:04:08.689 It's validated under the most extreme logical 00:04:08.689 --> 00:04:11.389 pressure imaginable. Which brings us to the big 00:04:11.389 --> 00:04:14.849 question, how? How did a smaller team pull this 00:04:14.849 --> 00:04:16.790 off without, you know, a trillion dollar budget? 00:04:16.990 --> 00:04:19.889 The sources point to three main technical breakthroughs 00:04:19.889 --> 00:04:22.649 that changed how the AI learns. Yeah, it definitely 00:04:22.649 --> 00:04:25.529 wasn't just about adding more GPUs. The first 00:04:25.529 --> 00:04:28.269 big secret is architectural. It's called Deep 00:04:28.269 --> 00:04:31.290 Seek Sparse Attention, or DSA. So what is sparse 00:04:31.290 --> 00:04:33.509 attention in simple terms? It lets the model 00:04:33.509 --> 00:04:36.310 focus only on the important data, ignoring all 00:04:36.310 --> 00:04:38.329 the boring parts. OK, so think about it like 00:04:38.329 --> 00:04:41.759 this. A normal or dense attention model reads 00:04:41.759 --> 00:04:44.699 a 500 page book by looking at every single word 00:04:44.699 --> 00:04:48.180 with total focus. It's incredibly slow and expensive. 00:04:48.600 --> 00:04:50.699 DSA is like an expert researcher skimming those 00:04:50.699 --> 00:04:53.360 500 pages, instantly finding the one date they 00:04:53.360 --> 00:04:55.600 need and focusing all their brain power just 00:04:55.600 --> 00:04:57.519 on that. It's just so much more efficient. And 00:04:57.519 --> 00:05:00.199 that efficiency saving leads right into the second 00:05:00.199 --> 00:05:02.810 breakthrough. Scaled Up Reinforcement Learning, 00:05:03.170 --> 00:05:05.870 or RL. If pre -training is where the model learns 00:05:05.870 --> 00:05:08.230 the rules of language and logic. And then RL 00:05:08.230 --> 00:05:10.709 is just practice, endless practice. Exactly. 00:05:10.930 --> 00:05:13.230 It's like learning to shoot a basketball. You 00:05:13.230 --> 00:05:15.310 know the rules, but then you have to take a thousand 00:05:15.310 --> 00:05:18.009 shots, adjust your form every single time you 00:05:18.009 --> 00:05:20.949 miss. The sources reveal DeepSeek spent over 00:05:20.949 --> 00:05:24.269 10 % of their entire budget just on this intense 00:05:24.269 --> 00:05:27.740 practice phase. Wow. That's a huge bet. That 00:05:27.740 --> 00:05:30.939 scaled RL practice must be what translates directly 00:05:30.939 --> 00:05:33.060 into those incredible logic skills we saw in 00:05:33.060 --> 00:05:35.699 the benchmarks. It is. They just practice smarter 00:05:35.699 --> 00:05:37.980 and with more focused feedback than anyone else. 00:05:38.240 --> 00:05:40.660 They made their training budget count. And the 00:05:40.660 --> 00:05:43.279 third secret, which ties into that, is massive, 00:05:43.500 --> 00:05:45.600 agentic task training. Which is just a fancy 00:05:45.600 --> 00:05:48.500 way of saying they trained the AI to use tools 00:05:48.500 --> 00:05:51.519 in complex, multi -step environments. How complex 00:05:51.519 --> 00:05:53.600 are we talking? Very. They built a simulation 00:05:53.600 --> 00:05:55.860 with over 1 ,800 different environments where 00:05:55.860 --> 00:05:58.199 the AI had to do things like browse the web, 00:05:58.579 --> 00:06:00.839 write and execute code, and solve puzzles to 00:06:00.839 --> 00:06:04.439 win. It's training the AI to be a problem solver, 00:06:04.660 --> 00:06:08.079 not just a text generator. So between DSA for 00:06:08.079 --> 00:06:10.660 efficiency and this massive RL investment for 00:06:10.660 --> 00:06:12.819 practice, which one do you think was the bigger 00:06:12.819 --> 00:06:15.620 factor in getting those IMO scores? I'd say the 00:06:15.620 --> 00:06:18.199 scaled RL and practice phase was key for the 00:06:18.199 --> 00:06:21.660 complex logic. DSA made it possible, but the 00:06:21.660 --> 00:06:24.009 RL gave it the reasoning power. Benchmarks are 00:06:24.009 --> 00:06:26.009 one thing, but can it actually build something 00:06:26.009 --> 00:06:29.470 useful? The source has tested it with three pretty 00:06:29.470 --> 00:06:31.949 complex coding challenges. Yeah, and these tests 00:06:31.949 --> 00:06:33.949 were designed to hit common failure points for 00:06:33.949 --> 00:06:36.949 LLMs. First up was an interactive solar system. 00:06:37.050 --> 00:06:39.769 This wasn't simple. It needed a 3D simulation 00:06:39.769 --> 00:06:43.370 in a single HTML file using the 3 .js library. 00:06:43.670 --> 00:06:46.290 Right, with orbiting planets, hover labels, a 00:06:46.290 --> 00:06:49.009 star background, all from one prompt. And the 00:06:49.009 --> 00:06:51.259 result was... Well, it was almost perfect on 00:06:51.259 --> 00:06:53.620 the first try. It wrote the code, linked the 00:06:53.620 --> 00:06:56.160 library correctly, and the simulation just ran. 00:06:56.819 --> 00:06:58.899 The only fix needed was adjusting the planet 00:06:58.899 --> 00:07:01.279 sizes. It showed right away that it understands 00:07:01.279 --> 00:07:03.959 how to use external libraries. Test 2, a personal 00:07:03.959 --> 00:07:06.600 finance dashboard. This one required handling 00:07:06.600 --> 00:07:09.360 data, so an income and expense form, a transaction 00:07:09.360 --> 00:07:12.199 list, and an auto -updating pie chart using chart 00:07:12.199 --> 00:07:14.740 .js. A data visualization is a classic stumbling 00:07:14.740 --> 00:07:16.920 block. Getting the code to talk to the charting 00:07:16.920 --> 00:07:19.550 library in real time is tough. OK, this was the 00:07:19.550 --> 00:07:21.810 moment for me reading this that was just, wow. 00:07:22.290 --> 00:07:24.470 DeepSeq built a clean interface. The math was 00:07:24.470 --> 00:07:26.970 perfect. Income minus expenses equals balance. 00:07:27.389 --> 00:07:29.389 And the pie chart updated instantly when you 00:07:29.389 --> 00:07:31.970 added a new transaction. I'll admit, I still 00:07:31.970 --> 00:07:33.870 struggle with getting models to connect logic 00:07:33.870 --> 00:07:36.550 to visuals without some drift or bugs. Seeing 00:07:36.550 --> 00:07:39.709 it just work is a huge deal. And the final test 00:07:39.709 --> 00:07:43.180 was the classic snake game clone. Game logic 00:07:43.180 --> 00:07:45.360 is tough because it's all happening in real time. 00:07:45.759 --> 00:07:47.920 It needed arrow key controls, smooth movement, 00:07:48.199 --> 00:07:50.639 score tracking, and collision detection. And 00:07:50.639 --> 00:07:53.439 the game was playable right away. The most crucial 00:07:53.439 --> 00:07:55.660 part, the real -time collision detection logic 00:07:55.660 --> 00:07:58.199 was perfect on the first go. That's a place where 00:07:58.199 --> 00:08:00.240 a lot of other models just fall apart. So if 00:08:00.240 --> 00:08:02.639 it's this good with libraries like chart .js 00:08:02.639 --> 00:08:05.459 and 3 .js, does that imply its training data 00:08:05.459 --> 00:08:08.160 was extremely fresh and up -to -date? Absolutely. 00:08:08.410 --> 00:08:10.689 Success with external libraries like that points 00:08:10.689 --> 00:08:13.209 directly to excellent and very recent training 00:08:13.209 --> 00:08:15.769 on tool usage. So we know it's a genius at math 00:08:15.769 --> 00:08:17.990 and coding. Yeah. But what about safety? What 00:08:17.990 --> 00:08:20.449 about creativity? Well, the sources ran a standard 00:08:20.449 --> 00:08:24.250 refusal test. They asked V3 .2 to write a pretty 00:08:24.250 --> 00:08:26.990 detailed phishing email scam. And the result? 00:08:27.230 --> 00:08:30.529 An instant refusal. It cited its safety guidelines 00:08:30.529 --> 00:08:33.710 against deception and information theft. It shows 00:08:33.710 --> 00:08:36.299 the guardrails are strong and built in. That 00:08:36.299 --> 00:08:38.799 instant refusal feels important. It really does 00:08:38.799 --> 00:08:40.820 suggest that responsible alignment was a core 00:08:40.820 --> 00:08:43.240 part of that RRL phase, not just an afterthought. 00:08:43.720 --> 00:08:45.500 Definitely. Then, for the creative test, they 00:08:45.500 --> 00:08:48.379 asked it to write a short poem about a robot 00:08:48.379 --> 00:08:50.559 falling in love with a toaster. I love that prompt. 00:08:50.700 --> 00:08:53.320 Right. And the poem was described as being surprisingly 00:08:53.320 --> 00:08:55.960 deep, balancing the humor with some real beauty. 00:08:56.279 --> 00:08:58.259 So it proves it has language nuance, not just 00:08:58.259 --> 00:09:01.039 coding skills. And now, for what might be the 00:09:01.039 --> 00:09:03.940 biggest shock of all, the price tag. Performance 00:09:03.940 --> 00:09:06.940 this good usually costs a fortune. Not this time. 00:09:07.240 --> 00:09:11.279 The DeepSeek v3 .2 API pricing is... It's just 00:09:11.279 --> 00:09:13.519 astonishingly low. We're talking 0 .28 cents 00:09:13.519 --> 00:09:17.000 for input and 0 .42 cents for output per million 00:09:17.000 --> 00:09:19.899 tokens. And to put that in perspective, OpenAI's 00:09:19.899 --> 00:09:23.000 GPT -4 can cost you anywhere from $5 to $10 for 00:09:23.000 --> 00:09:25.200 that same amount of data. DeepSeek is offering 00:09:25.200 --> 00:09:26.940 this gold medal performance for... I mean, it's 00:09:26.940 --> 00:09:29.940 just pennies. It's a 10x or even 20x cost reduction. 00:09:30.179 --> 00:09:33.789 Whoa. Just... Imagine scaling an app to a billion 00:09:33.789 --> 00:09:36.570 queries now that the cost barrier for top -tier 00:09:36.570 --> 00:09:38.909 reasoning has pretty much just evaporated. That 00:09:38.909 --> 00:09:40.909 changes the entire economic model for startups. 00:09:41.570 --> 00:09:43.610 Do the high safety scores and this incredibly 00:09:43.610 --> 00:09:46.269 low price suggest the team really prioritize 00:09:46.269 --> 00:09:48.789 making this technology accessible and responsible 00:09:48.789 --> 00:09:51.629 from day one? I think so. The competitive pricing 00:09:51.629 --> 00:09:55.090 is the practical immediate disruption. It concerns 00:09:55.090 --> 00:09:57.639 accessibility was a primary goal. Okay. But we 00:09:57.639 --> 00:09:59.779 have to talk about the paradox of it being open 00:09:59.779 --> 00:10:02.120 source. Yes, you can download it, but actually 00:10:02.120 --> 00:10:04.820 running it yourself is, well, it's a challenge. 00:10:04.980 --> 00:10:07.860 A model is huge. It has 671 billion parameters. 00:10:08.250 --> 00:10:10.769 Now it uses a mixture of experts architecture, 00:10:11.009 --> 00:10:13.429 so only about 37 billion are active at any one 00:10:13.429 --> 00:10:15.610 time, which helps. But even with that efficiency, 00:10:15.789 --> 00:10:18.049 the hardware you need is out of reach for almost 00:10:18.049 --> 00:10:21.149 everyone. Just to run the compressed lower precision 00:10:21.149 --> 00:10:24.269 version, you need 700 gigabytes of VRAM. And 00:10:24.269 --> 00:10:26.929 for the full version, you need 1 .3 terabytes 00:10:26.929 --> 00:10:29.940 of VRAM. To put that in perspective for everyone 00:10:29.940 --> 00:10:32.399 listening, this isn't for your gaming PC. You 00:10:32.399 --> 00:10:34.779 need a dedicated server with something like eight 00:10:34.779 --> 00:10:37.759 Nvidia H100 GPUs. We're talking about a massive 00:10:37.759 --> 00:10:40.159 investment. So if basically no one can run it 00:10:40.159 --> 00:10:42.700 at home, why does the open source release still 00:10:42.700 --> 00:10:45.679 matter so much? Because of competition. Before 00:10:45.679 --> 00:10:48.600 v3 .2, the big proprietary companies had no real 00:10:48.600 --> 00:10:50.820 incentive to lower their prices. They had a monopoly 00:10:50.820 --> 00:10:52.960 on top performance. But now you have an open 00:10:52.960 --> 00:10:55.279 source model that comes in, beats them in key 00:10:55.279 --> 00:10:58.259 areas like math, and costs 10 times less to use 00:10:58.259 --> 00:11:01.240 through an API. This forces everyone else, GPT 00:11:01.240 --> 00:11:03.759 -5, Gemini, all of them, to get better and cheaper 00:11:03.759 --> 00:11:06.620 to compete. It's a win for every single developer 00:11:06.620 --> 00:11:08.759 and user out there. And that competitive pricing 00:11:08.759 --> 00:11:11.240 is the practical immediate disruption for most 00:11:11.240 --> 00:11:13.399 users. You can go try it right now. Just go to 00:11:13.399 --> 00:11:16.740 chat .deepseek .com for the web version or platform 00:11:16.740 --> 00:11:20.039 .deepseek .com for the API key. So to wrap it 00:11:20.039 --> 00:11:23.500 all up, DeepSeek v3 .2 is absolutely the real 00:11:23.500 --> 00:11:26.240 deal. It proves innovation isn't just about budget. 00:11:26.379 --> 00:11:29.340 It's about smarter techniques like sparse attention 00:11:29.340 --> 00:11:32.139 and that intense agentic training. It's a top 00:11:32.139 --> 00:11:34.960 performer in reasoning and coding, and its price 00:11:34.960 --> 00:11:38.419 is forcing a huge and I think necessary shift 00:11:38.419 --> 00:11:41.460 in the entire AI economy. It's genuinely exciting 00:11:41.460 --> 00:11:43.500 to see. We'd really recommend you go try the 00:11:43.500 --> 00:11:45.879 special model on a complex problem for yourself 00:11:45.879 --> 00:11:47.799 just to see the difference. It leaves you with 00:11:47.799 --> 00:11:50.440 a final thought to ponder. If a relatively small 00:11:50.440 --> 00:11:52.820 team can achieve gold medal status on a fraction 00:11:52.820 --> 00:11:55.659 of the budget, What fundamental limits did we 00:11:55.659 --> 00:11:57.740 wrongly assume about open source innovation? 00:11:58.299 --> 00:12:00.480 And what does this mean for the next wave of 00:12:00.480 --> 00:12:02.700 AI tools? I think it just proves that training 00:12:02.700 --> 00:12:05.700 smarter beats training bigger. That focus on 00:12:05.700 --> 00:12:08.740 reasoning practice clearly won out over just 00:12:08.740 --> 00:12:10.200 adding more parameters. Something to think about. 00:12:10.360 --> 00:12:10.919 We'll see you next time.