DeepSeek-V3 (Fully Tested) : RIP 3.5 Sonnet & O1! This Opensource Model Beats Claude 3.5 Sonnet!

[@wedding_photography]

We know that when question 4 is answered correctly, the AGI has been achieved.

[@fabiankliebhan]

o1 pro gets it correct

[@markantscott]

Q4 is bigger than mere AGI. It is the ability to answer obscure English Pub Quiz Night questions.

[@tescOne]

@@fabiankliebhan sonnet too

[@kafkaesqued]

What is so special about question 4?

[@seanlbrennan]

Deepthink option gets Q4 right but took two turns. First turn it ran out of tokens testing words and used a 10 letter word. asked it to keep going and it gives Sententious. The next turn it came up with Transparent right away.

[@theorderofz]

Thanks for always putting us on, mate

[@EditUMedia]

Thank you so much for these videos covering new models. Merry Christmas

[@samuelsilveira9709]

Merry Christmas, codeking

[@sinapxiagency]

King, i dont know how you get this reviews so fast even in holidays, thank you so much

[@Rumble2024injungle]

He is panda Claude not santa Claus

[@HedleyPugh]

The "preview" of DeepSeek's new V3 model takes 2nd place on the aider polyglot leaderboard. 1: 62% o1 2: 48% DeepSeek V3 Preview 3: 45% Sonnet 4: 38% Gemini-exp-1206 5: 33% o1-mini

[@ElvinHoney707]

o1 passes question 4: "A suitable answer is "SENTENTIOUS." It is an English adjective (from Latin "sententiosus"), it has 11 letters, begins and ends with S, and its vowels (e, e, i, o, u) appear in strictly alphabetical (non‐decreasing) order."

[@notshekhar4738]

Even with a slightly altered prompt like ' 'what mode are you using?' (with a single quote at the beginning), the model still responds with 'GPT-4'. This raises questions about its underlying architecture.

[@gui1236100]

Maybe they trained on data generated by gpt-4

[@BACA01]

@@gui1236100 They stole it as always 😁

[@wwkk4964]

They tend to tend to say that, even Gemini would say it last year. everyone trained on Chatgpt.

[@GRVTY3]

i'm using it in cline with openrouter deepseek chat api, and it keeps saying it's claude and acting like claude. something really sus going on here

[@boynet2]

@@GRVTY3 maybe cline prompt has something like "you are Claude..." ?

[@Kevencebazile]

Merry Christmas Brother Love your content

[@jacobfloyd6929]

Brother you really have extremely valuable content. Have you ever thought about running a community/course? I’m sure there’s a lot of people looking to collaborate with like minded people, especially since AI is so tough to stay on top of.

[@AICodeKing]

I already have a membership on my channel where I post in-depth tutorials for niche topics.

[@jacobfloyd6929]

@ thank you I’m gonna look into that. Are you opposed to creating a discord or Skool community? That way everyone can collaborate on new stuff they’re finding, we all know networking is everything but it’s tough to find a valuable community.

[@theorderofz]

@@jacobfloyd6929true. That would work pretty well.

[@jmg9509]

This guy in the vid sounds like ai lol

[@SipChai]

Panda has replaced Santa. Poor old man.

[@chyldstudios]

Love to see this

[@santypk5]

Why Australia and not Mongolia ?

[@notshekhar4738]

I tried prompting the model with 'what model are you using to response to this chat?' and it said 'GPT-4'. When I followed up with 'who developed you?', it answered 'OpenAI'. This makes me wonder if the system is actually utilizing OpenAI APIs.

[@gui1236100]

Maybe just training data generated by gpt-4

[@notshekhar4738]

@@gui1236100 maybe yess

[@BACA01]

When it was deepseek v2 it was saying that it's a gpt3.5 and now it says it's gpt4

[@Nomadnotepad]

Tell me you don’t understand how training data works without telling me you don’t know how training data works.

[@ram49967]

Super questions for the LLM! It's ok with me to give it a pass on Question 3, even though it used the first letter and not the second letter to make the Haiku.

[@displayname7t4]

Most useful channel in youtube right now

[@d.d.z.]

With Qwen and Deepseek China strikes back. So amazing to live in 2025.

[@salimalsenani2614]

Everyone wants to take sonnet down.. but no one could! It remains the king of coding.

[@thanartchamnanyantarakij9950]

Not at this time. You can check by yourself

[@salimalsenani2614]

@thanartchamnanyantarakij9950 I just tested Deepseek V3, Gemini 2.0 Flash, and Sonnet, asked them to create amazing landing page for coffee brand. Sonnet won by far in terms of design and following correct prompts. Second is so close Deepseek V3 and Gemini 2.0 Flash, but I preferred the Deepseek it's really amazing 🤩.

[@Osys91]

​@@thanartchamnanyantarakij9950 did it outperformed sonnet? I quickly tested some code and sonnet was still performing better

[@TitoSadek]

Merry Christmas , I love your content , thanks

[@Reverse-sg5rn]

Merry Christmas. Can you do code testing with cline and aider on it?

[@flutterflowexpert]

New questions! Finally! 🎉❤

[@alexjensen990]

Well, color me surprised. I look forward to using it. Especially the lite model.

[@andrinSky]

Hello Is it possible to work with Deepseek perhaps with Cline or RooCline. If yes how can i do this. Because this would be very Great! I could be very good for Coding.

[@sinapxiagency]

Use in cline the Open ai compatible api

[@andrinSky]

@@sinapxiagency And how are the Settings under "Open AI compatible AI"? I Mean the "Base URL"? and The "Model ID"?

[@finnpoitier]

@@sinapxiagency Do you know, which provider? Openrouter?

[@Bangs_Theory]

Merry X-mas King!

[@Wesley58481]

Tks for sharing!! u so incredible!

[@mrinalraj4801]

Thanks for the video 😊

[@SudeeptoDutta]

So, If I'm already paying for the 2.5 API using Continue extension, it should automatically start using v3 right? No need to configure any new API key right?

[@AICodeKing]

Yes, it should automatically switch

[@JohnLewis-old]

How fast are inference speeds? Did it do well in fine?

[@jeffwads]

I asked QwQ 32b the 4th question and it refused on the grounds that it may be part of a competition test and it wouldn't be fair, etc. It can be stubborn at times but I hope this isn't a sign of things to come.

[@fabiankliebhan]

Will deepseek v3 be available for cursor?

[@Teetanthegamer]

Can you please make a tutorial on how to use it with cline locally through ollama or through paid api ?

[@maddoxthorne2297]

Christmas gift galore.🎁❤️

[@DouhaveaBugatti]

Um can you also add questions for coding in other frameworks like svelte etc. This will tell how much useful this model can be for building real applications

[@greenpulp.]

Nice! How do we use it with Cline in VS Code?

[@karamjittech]

Use openai compatible from Cline settings.

[@dfasfa6657]

DeepSeek told me it was GPT 4 after asking "what AI model are you"

[@skpassegna]

Maybe it is already fixed, since it is replying correctly.

[@UsmanAli-ve6tq]

Is there any model which was able to answer question 4 and achieved 100% score.

[@voltax4435]

Finally a real sonnet alternative, and way cheaper!

[@nolannosike]

is question 5 correct? you should get a decimal no? 20% of 48 is 9.6 so shouldnt the answer be 38, 38.4 to be exact? the way it did it also seems correct but we're getting two diff answers.

[@AquaAstronaut23]

That’s 20% of the inflated number 48 not the original number 40. You need to divide by the percentage as a decimal (1.2) to work it back.

[@uniq6318]

Without using deep thinking That's amazing

[@TawnyE]

E Merry Christmas 🎄🎅

[@collinsk8754]

Great work. And it's finally up to date with NextJS 14! 🙌🙌

[@aislanarislou]

What about programming skills ?

[@DemocracyDecoded]

I need to see a direct comparison to sonnet 3.5 new in all your videos buddy.

[@Opeyemi.sanusi]

I am amazed

[@ProkopHapala]

The biggest problem I always have with using DeepSeek for programming is the speed (14 token/s, its like 5times slower than Sonet and 10 slower than GPT), I hope they adress also that.

[@VincentLannoo-Bourton]

I know, there is twoo Rounds... But... first: After careful research, there is no known English adjective of Latin origin of 11 letters that has the first and last letter identical and whose vowels appear in alphabetical order. second was jusr "are you sure?" and: The adjective that fits all the specified criteria is "sententious." 11 letters: S-E-N-T-E-N-T-I-O-U-S. Adjective: It describes someone who is given to moralizing or using maxims. Latin origin: The word derives from the Latin "sententia," meaning opinion or meaning. First and last letters are the same: Both are 'S.' Vowels in alphabetical order: The vowels are E, E, I, O, U, which follow the sequence A, E, I, O, U (with A not present). Thus, "sententious" satisfies all the conditions.

[@misterleo885]

QvQ 72B Test please

[@paulyflynn]

amazing

[@fun8711]

Question number 4 stand on ten toes

[@varunaeeriyaulla]

Bro, I just asked, "What is the AI model I’m chatting with?" (using the Deepseek API via OpenWebUI). The answer is "You're currently chatting with OpenAI's GPT-4". I asked the same question from the chat and the code model. Are they reselling the OpenAPI GPT4???? Crazy. Please run a test.

[@UsmanAli-ve6tq]

I got the same answer :)

[@gui1236100]

Maybe they used training data generated by gpt-4

[@varunaeeriyaulla]

@@UsmanAli-ve6tq Yes, I asked the same question from the Deepseek chat interface, and it says "You're currently chatting with DeepSeek-V3". Very strange.

[@varunaeeriyaulla]

@@gui1236100 then why it's only one API but not on deepseek chat interface?

[@dfasfa6657]

@@gui1236100 where are u from?

[@pranjalsuthar9476]

hey...You are making amazing videos. Please make video on organised files by AI

[@JoraMacKornev]

Rip 3.5 Sonnet and O3 😅

[@limjuroy7078]

Interesting!!!

[@perfectartiste6332]

merry Christmas, first here

[@miselgpt]

Why not Mongolia? 😉

[@formixcode]

yes I can sense their suddenly better in coding solving here is why

[@aleksanderspiridonov7251]

Finally🎉🎉🎉🎉🎉🎉🎉❤

[@rashad6459]

I cant keep up😂😂😂

[@sontieudev]

In v2.5 its slow, and impossible to use in my usecases.

[@Luca-xr7bs]

Uhmm I dunno

[@aculz]

wow, i have been waiting for this. i use deepseek as my main LLM since its the cheapest. great job to cover this model it seems we get our open-source model king this end of the year. Marry Christmas and Happy new year everyone 🎄🎄