🚀 Add to Chrome – It’s Free - YouTube Summarizer
Category: Tech Review
Tags: AICodingDeepseekModelTech
Entities: ADER leaderboardsDeepseek V3.1Microsass FastRequestySonnetTMDB API
00:00
[Music] [Applause] Hi, welcome to another video. So, Deepseek V3.1 was launched, but the major info about it wasn't available before.
Now, it is available, and the
00:17
TLDDR is that it is a hybrid reasoning model that can basically do both reasoning and non-reasoning tasks. This allows for efficient performance.
It also now has much better tool calling support which was one of the major pain
00:34
points of Deepseek R1 and V3. It now scores way better in multiple tool calling benchmarks and works much better in things like clin rue and open code.
Similarly, they have also increased the
00:49
context limit to 128K from 65K which is a great upgrade as well. The model endpoints still exist the same way which is deepseek chat for the general model and the reasoning endpoint
01:05
for the reasoning variant instead of anything like reasoning effort. The performance is really improved here on my benchmark.
The model is way better now even without reasoning. It can create floor plans pretty well though
01:21
they are not the best. The oneshot generation that just works for a model this cheap is awesome.
The SVG Panda with the burger is one of the best I have seen yet. It is actually usable, which is not even something that Opus
01:36
can make. The autoplay chess also works pretty well.
It doesn't always make legal moves, but it is still better than many other generations. The flying butterfly is also kind of cool.
And overall this stuff is really
01:52
good. If we talk about the reasoning variant then it is also good but not as good as the non-reasoning variant at least for coding.
So yeah it's not as great at least for coding but the non-reasoning model is
02:08
actually amazing in ADER leaderboards. It also comes close to opus for way cheaper pricing which is quite awesome for an open and cheaper model.
It's really good and I think it is now one of the best options
02:24
to be a drop-in replacement for sonnet and that's what DeepS also thinks because they themselves now provide an anthropic API format along with details to make it a drop-in replacement in claw code and use this model. I've been using
02:42
it and it is really awesome. It feels a lot like Sonnet but it's 20x cheaper.
So yeah, that's a great deal. Let me show you how you can also do the same and use it in interesting ways.
But before we do
02:57
that, let me tell you about today's sponsor, Microsass Fast. Dreaming of launching a Micros or AI side project, but wasting weeks setting up O payments and SEO?
Check out Microsass Fast, a next.js boilerplate with clerk, stripe, resend, hostgsql, and AI instructions
03:15
that cut hallucinations by 90% for vibe coding. Easy back-end integration with Python, Node, and Go.
It is built and used by a CTO who helped 50 plus founders to launch SAS in the past year. You can save 50 plus hours and actually ship faster.
Check now. Link is in the
03:31
description. Now, back to the video.
First, make sure that you have Claude code installed and upgraded to the latest version. You can do either of them with the same command.
Once that is done, you can export the
03:47
environment variables for base URL as Deepseek's anthropic base URL along with O token model and also the small fast model as deepseeek chat model. Also, open router doesn't support
04:03
anthropic format. So you can use claude code router for that or a better option that I use is requesty as it is way better than open router in terms of configuration and customization.
It also comes with anthropic compatible
04:21
format and you should be able to use almost any model from there. I am going to be using deepsek from there but you can use the original API as well.
For Requesty, you just have to change the base URL to support Requesty as
04:38
written in their docs and then just add deepsee slash in front of the DeepSeek chat endpoint and it should work. It actually works pretty well.
The slash commands here all work fine along with MCPS which also work fine. For example,
04:57
I use the bite rover MCP here because I really like it for a memory layer. It allows you to plug it into all kinds of coders and it can create memories that are synced across all your coders.
You can also share it with your teammates
05:12
which allows you to build rules and memories that stick to your projects. So yeah, this is something that I use and you can easily get it configured on cloud code.
Just do that and you'll get some better results.
05:28
For example, here I have it started in a Nex.js project. I'm going to ask it to make me a simple movie tracker app that uses the TMDB API for the database.
And what you'll see is that it actually goes ahead and works amazingly well. It can
05:46
do most of the tool calling without failing, which was a big downside in the previous version, but now it is amazingly better and actually works on par with Sonnet while being 20 times cheaper, literally. So yeah, it's
06:02
awesome. It can also use the to-do list stuff as well as MCPs without any issues.
It's really great. for planning.
You can change the model to Deepseek's reasoning variant as well if you wish to do that.
06:18
Anyway, the speed of the model is also now way better. It's probably because they don't need to deploy two models anymore and just one works, which produces way better speeds now and works well.
So, it's pretty usable now.
06:35
Anyway, in a bit it gets it done. And you can see that it did this well without any major errors.
And we can now go ahead and use it. It looks pretty great.
It's on par with what Sonnet makes in one shot while being a lot
06:51
cheaper and pretty similar. So yeah, this makes a lot of sense to use.
I prefer it over GLM for replacing and clawed as it's pretty great now. I was thinking Deepseek V3 3.1 was not as good at first, but as I used it, it felt like
07:08
Deepseek actually addressed almost all the major pain points of the original model, like tool calling, MCP support, and everything. It now works way better and is also way better at coding.
So yeah, I'd highly recommend you check
07:25
this out. It's an awesome model.
I don't know what happened with the reasoning variant. They probably just have that for things like mathematics because the non-reasoning variant now works amazingly well for coding and there's
07:40
probably no need to use sonnet for trivial tasks which saves a ton of money and costs. I really like it.
Go ahead and check it out and use it all you want. Overall, it's pretty cool.
Anyway, share your thoughts below and subscribe
07:56
to the channel. You can also donate via super thanks option or join the channel as well and get some perks.
I'll see you in the next video. Bye.
[Music]