Speaker 1: A huge release out of DeepSeek today. So DeepSeek R1, their reasoning model, is now fully open source with an MIT license. What's interesting with this model is it is the first major competitor to OpenAI's O1 series of models. In this video, I'll give you a brief overview about the model, the pricing, some of the technical details, and then I'll show you how to get started with the model. Right off the bat to go over some aspects about this model. It's a mixture of experts model in terms of the active parameters. It's a 37 billion active parameters. What's interesting with this is it's actually the same size as DeepSeek V3. For the MMLU score, it scores a 90.8. In comparison, CLAW 3.5 Sonnet has 88.3, GPD 4.0 has 87.2, respectively. And this even outperforms O1 Mini, and it is just shy of OpenAI's O1 that was just released last month. Just about all of the metrics for the coding benchmarks, with the exception of Adar Polyglot, are just shy of O1. But where you can really see the difference on some of these benchmarks is look at Codeforces, for instance. CLAW 3.5 Sonnet scores 20.3, GPD 4.0 23.6, R1 scores 96.3. Another huge thing with this release is it has a very permissive license. It's a MIT license. You can distill the model, you can commercialize the model, you can use the outputs to fine tune other models, or use the outputs from the model for synthetic data generation, or basically whatever you want. They also released six small models that are fully open source as well. These are open source distilled models from R1. We can see here DeepSeek R1, we have 1.5, 70B, 14B, all the way through to 70B, and the 32B and 70B, these are on par with OpenAI's O1 Mini. What's great with these distilled models is you're going to be able to run this basically regardless of the hardware that you have. Now to quickly touch on the technical report, I'll also link everything within the description of the video if you're interested. We introduce our first generation reasoning models, DeepSeek R10 and DeepSeek R1. DeepSeek R10, a model trained via large-scale reinforcement learning without supervised fine tuning, as a preliminary step demonstrating remarkable reasoning capabilities. Through RL, DeepSeek R10 naturally emerges with numerous powerful and intriguing reasoning behaviors. However, it encounters challenges such as poor readability and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek R1, which incorporates multi-stage training and cold start data before RL. DeepSeek R1 achieves performance comparable to OpenAI's O1-1217 on reasoning tasks. To support the research community, we open source DeepSeek R10, DeepSeek R1 1.5b, 7b, 8b, 14b, 32b, 70b, which were all distilled from DeepSeek R1 based on Kuen and Lama. If you're interested in diving in a little further, I'll link everything within this technical report within the description of the video. To try out the model, you can try it out at chat.deepseek.com. What's really great with this is you also see the reasoning steps here, which is different than what you get from OpenAI's O1, where you don't actually see. Here we see our short story, but now if we actually test some of the reasoning capability, if I say how many times did the letter R occur in that story, I'll put that through and we see okay, the user wants to know how many times the letter R appears in the story. What it's doing is it's breaking up all of the sentences and then with each sentence, it's breaking up every single word and it's counting word by word. At the bottom, we have the breakdown of all of the different paragraphs and potentially how many times the letter R occurred within that. Now, in addition to the chat interface, you can also access this from their API with the model string of DeepSeek Reasoner. Next, what's really unreal with this model is the pricing. For a million tokens per input, it is 55 cents. With context caching, it's 14 cents. And then for the output tokens, it is $2.19. So if we compare that to O1, output tokens are $60, a million tokens of input is $15 and context caching is still $7 and 50 cents per million tokens. The pricing alone from their hosted API makes this model incredibly interesting in terms of some of the specifics of their API defaults to 4000 tokens of output with a maximum of 8000 tokens. But with a chain of thought the output can be as high as 32,000 tokens, there isn't the ability to control the reasoning effort right now like you can with the O1 models. What I love with the output is you can get both the reasoning content, so all those steps, and then parsed out individually in a separate key is the actual content of the response as well. Aravind Srinivas mentioned that DeepSeek has largely replicated O1 mini and has open sourced it. From Technium put it well, it destroys 4.0, we have frontier models at home, I guess now. Another thing worth mentioning is it's free to use on the web page what I demonstrated for you with the exception of the API that is all free. And you can also download a mobile app that they just released. The last I just want to show you how you can set this up within your IDE if you're interested. So continue is a tool that I love, you can use this within VS Code, Windsurf, or Cursor. And what you'll be able to do is once you've installed continue from the extension marketplace, you can go to the bottom here and add in a new chat model. What you can do here is even if you don't see the reasoning model from the drop down, what you can do is within the configuration file is if you just set the model string to DeepSeekReasoner, you set the context length, and you put in your API key, you'll be able to use this within your IDE. If I just demonstrated here, if I just say write me a proxy server for opening eyes chat completion endpoint, pretty quickly, we see the responses here, it's streaming all back. And you can use it similar to something like you would cursor, I can just add that all in here. And we have what looks to be like a great response. So this is a really good option. If you're looking to use this in the context of an IDE. And the thing that I noticed with this model, especially from their API is what is very, very quick. Overall, that's pretty much it for this video. Kudos to the team over at DeepSeek on this release, it is really great to be able to have these types of models out there within the open source community. Let me know your thoughts in the comments below. And if you found this video useful, please like, comment, share and subscribe. Otherwise, until the next one. Transcribed by https://otter.ai