DeepCoder-14B: NEW Opensource Coding Model Beats 03-Mini! (Tested)

Looks like we have a brand new fully open-source AI coding model that was just released today by Together AI and it's a gamecher. Allow me to introduce Deepcoder, a 14 billion parameter model that astonishingly matches the performance of Open AI's 03 Mini. It's built in collaboration with Aentica. This is where they've released Deepcoder 14B preview, which was trained through distributed reinforcement learning on 24K verifiable coding problems over just 2. 5 weeks using 32 H00 GPUs, which is kind of insane. It hits 60. 6%age 6 percentage pass rate on at one on the live benchmark test at an 8% boost over its base and it ranks in the 95th. 3 percentile on code forces with support for only 64 token context. It is something that's kind of lightweight but it is still fully open source and it reaches these high levels as 03 mini. The model weight, the data set, the training log as well as a lightning fast pipeline called world pipe is fully open source. Deepcoder marks a huge leap forward for the open source community and this is something that we're going to be exploring throughout today's video. Now, I'll leave a link to together AI's blog post on Deepcoder cuz it is quite interesting as to how they developed this model. For example, to train deep coder, the team actually curated a highquality data set that contained 24K verified coding problems from trusted sources and they filtered out easy duplicate or broken problems to ensure stable and reliable training. The model was also trained in an isolated code sandbox environment so it would be capable of running thousands of unit tests per batch to accurately evaluate its output. A strict reward system was also used so that the code that was passed through all the tests would earn a credit which would help prevent the model from learning shortcuts. They've also used the smarter training algorithm and gradually increase the model's context length which was enabling it to handle inputs up to 64k tokens. I know it doesn't seem a lot but it's still decent for a small parameter size model. And to speed things up, they also introduced a new system optimization that cuts training time in half. And from the benchmark scores, you can see that the Deep Coder does quite well in comparison to these larger models. You can see that in comparison to 03 Mini, 01, Deep Seek R1, as well as Llama 4 Behemoth, it even beats it. But in other cases, it does quite well and it stands its ground against these proprietary models except DeepC because they're obviously an open- source firm. But you can see that it is holding its ground as a small parameter size model and it is definitely doing a great job in comparison to many of these other models. This is why today we're going to be testing this model out and showcasing if it is worth actually downloading or not. Now, if you are wanting to get started with this locally, you can go over to HuggingFace, copy the model card and install this with something like LM Studio or with a chatbased UI system like open web UI. Or you can go ahead and install it with Olama. And you can easily do this by going ahead and installing the 14B model by running this command within your command prompt. And obviously, you need to make sure that you have Olama installed beforehand and running in the background. But say if you want to just try it out to see if this is something that you would want to install, you can use glff. hat, which is a way for you to run any open source model for free. I believe they have $10 credits that you can use whenever you create an account. And this way you can experiment with the model to see if it's actually worth it. So, if you want, what you can do is copy the hugging face model card link for Deepcoder 14B and then you can head over to GHF paste in the link and you can then select the model that you want to work with. So, Deepcoder 14B and then you can go ahead and click launch and then now you will be able to start interacting and chatting with this model within this interface. So, let's just get started. We're going to start off with something super simple. Creating a simple CRM dashboard app for me. And you can see right away it is going to go ahead and start generating the code. So now first it is going to create the project structure. And then it's necessary code that is needed for this app to be functional. So once it has finished generating everything, I'm going to go ahead and paste it into a file and open it up. And there we go. We finally have all the different sorts of files outputed into the main CRM dashboard folder. And this was what the 14 billion parameter model was capable of outputting, which is definitely quite impressive. A 14 billion parameter model outputting a front end for a CRM dashboard. Now, I simply clicked on add customer. And let's see if you're capable of adding things. And it actually works, which is kind of

impressive. I thought it wouldn't work. And it does look like it shows the total customers. Now, if you go over to the customers tab, that was something that it didn't generate, which is kind of obvious with the 64K context window. But overall, I'm definitely impressed to see that it was capable of doing this. Now, next up, I'm going to have it send over a hard prompt to generate an SVG code for a symmetrical, elegant butterfly illustration. Now, as you all know, this is a prompt that many models tend to fail at. There's only a handful of models like the new Gemini 2. 5 Pro. Couple of the different OpenAI models, but some of them don't even generate it perfectly. Deepcar1 does. But you can see that this is a really difficult prompt. And to be honest, I'm kind of wondering if this model can actually output this based off the first performance I saw from it. So, let's go ahead and copy this and paste it into an SVG online viewer. So, I've copied the code and we're going to go ahead and paste it in here. And this is how it looks. So, I guess it got the antenna right. It got I guess the eyes right, the body is decent. It's just the wings that aren't attached as well as everything else. So, it did fail, but I'm definitely surprised to see that it did kind of get a bit of it right. Now I'll just simply go back into the chat uh chat interface and ask it could you please improve the design because it doesn't look anything like a butterfly. So one thing you'll notice whenever you're chatting with this model is that it talks to itself and you can see that it is working on trying to find a solution for this and now it looks like it has generated a new code snippet for this butterfly representation. So hopefully this looks a lot better. I'm not expecting it to generate it. And it still looks pretty much the same as the last one. It's just the wings got closer, I guess. But overall, this is definitely kind of unfortunate, but as expected for a 14 blade parameter model to not actually complete this task. In this case, I'm requesting it to do another task in SVG, and I'm asking it to create a smiley face. And it does get the job done. I tested it out, which is pretty great to see. But in terms of generating complicated SVG code, it can actually do this. But overall, it's pretty decent in most of the cases with front-end development as well as creating functional uh dashboards with different tools. Now, what I also noticed is that this model does a decent job in debugging different code snippets. This is where I have provided faulty code and I'm asking it just to simply fix the faulty code without identifying what's wrong. In this case, you can see there's a missing self in method definition. There's also a couple of other errors like uh not having or having the wrong data type for age. Uh the greeting method doesn't return anything. And you can see that it reasons to itself, works on finding a solution and at the end if you keep scrolling and this is basically the output I got. It is technically correct where the code will run based off of the fixes that it made. But the only thing that is incorrect is the string over here which says 25 when it should have been a number value as 25. And you can see that overall the output would be great and it did fix the different definitions that were faulty. And this is why I really think that this could be a great model that you could use if you do not have the resources to run bigger or larger models on your local computer. Now, at the end of the day, all of these different benchmark scores always show that it's the best model whenever a new model drops. But what I definitely recommend that you do is try it out yourself to see if it is even worth using or to get a better understanding if it is actually what it actually holds up in the benchmark scores. This is why I never trust benchmark scores and usually just test it out on my own to see if this is something that is capable and worth using. But overall, I think that this is a great 14 billion parameter size model. The only negative is the context window. But aside from that, it is open source. Everything is uh available for you to access the open weights as well as the training and anything that you see over here. It's fully accessible. So that is the plus side. But I hope you enjoyed today's video, guys, and got some sort of value. Let me know what you guys think about Deep Coder. I'll leave all these links in the description below. Follow me on the newsletter, the Discord, make sure you follow me on Twitter, and lastly, make sure you guys subscribe, turn on notification bell, like this video, and please take a look at our previous videos because there's a lot of content that you'll truly benefit from. But with that thought, guys, have an amazing day, spread positivity, and I'll see you guys fairly shortly. He suffers.