DeepSeekreleased its V3 modellast month . The company has nowunveiled its reasoning model , DeepSeek R1.DeepSeek claim it not only matches OpenAI ’s o1 good example but also outperforms it , particularly in math - related questions . The good thing is that an R1 model is open - source , free to expend , and can even extend topically . rent ’s explore if R1 is really that good .
Table of content
What is DeepSeek R1?
DeepSeek R1 is a abstract thought fashion model , meaning it does n’t just provide the first solution it finds . Instead , it “ consider ” through problems step by step , taking moment or even minutes to reach a answer . This deliberate chain - of - cerebration outgrowth shit it far more accurate than traditional AI models and in particular useful in areas like maths , physic , and coding , where logical thinking is of the essence .
DeepSeek achieves this reasoning capableness through a combination ofReinforcement Learning ( RL)andSupervised Fine - Tuning ( SFT ) . What ? Here ’s what these two terminal figure mean :
Initially , DeepSeek trust solely on Reinforcement Learning without fine - tuning . This “ DeepSeek R1 Zero ” phase demonstrated impressive reasoning abilities , including ego - check , reflection , and generating long chains of thought . However , it face up challenge such as pitiful legibility , repeating , and language mixing . To speak these issues , DeepSeek combine RL with Supervised Fine - Tuning . This threefold glide slope start the mannikin to refine its reasoning , memorize from preceding mistakes , and save systematically best final result . More significantly , this is an open - generator model under theMIT License .
The Numbers Behind DeepSeek R1
DeepSeek R1 boast a massive671 billion parameters . call back of parameters as the brain cellular phone an AI uses to learn from its training data . The more parameters a model has , the more elaborate and nuanced its intellect . To put this into perspective , while OpenAI has n’t unwrap the parameters for o1 , expert estimate it at around200 billion , making R1 significantly larger and potentially more powerful .
Despite its size , R1 only activates37 billion parameters per tokenduring processing . DeepSeek says it is done to guarantee the manikin remains efficient without compromise abstract thought capabilities .
The R1 model is built with the DeepSeek V3 example as its floor , so the computer architecture and other stats are mostly like . Here are the DeepSeek R1 framework stats :
How Does R1 Compare to OpenAI’s o1?
When it comes to benchmark , DeepSeek R1 is on equation with OpenAI ’s o1 modeling and even slimly surpasses it in domain like math . On math benchmarks like AIME , it score 79.8 % , slightly better than o1 ’s 79.2 % . For programming tasks on Codeforces , it outperformed 96.3 % of human programmers , showing it ’s a serious competitor . However , it ’s slightly behind o1 in coding benchmarks .
For developers , the model is cheaper to mix into apps . While the o1 example costs $ 15 per million input signal token and $ 60 per million outturn tokens , R1 costs just $ 0.14 per million input tokens ( Cache Hit ) , $ 0.55 for million input item ( Cache Miss ) and $ 2.19 for output token , making it 90%-95 % cheaper .
Another standout feature article of R1 is that it shows itsentire thought processduring reasoning , unlike o1 , which is often faint about how it arrives at solution .
Distilled Versions for Local Use
DeepSeek has also releaseddistilled modelsranging from1.5 billion to 70 billion parameters . These smaller fashion model hold back much of R1 ’s reasoning power but are lightweight enough to fly the coop even on a laptop computer .
Distilled Models:
These small models make it easy to test advanced AI capabilities topically without take expensive servers . For representative , 1.5B and 7B models can function on laptop . Whereas , 32B and 70B models deliver near R1 - tier public presentation but ask more potent setups . Even better , some of these model outgo OpenAI ’s o1 - miniskirt on benchmarks .
Also Read :
How to Access DeepSeek R1
DeepSeek R1 is easy to get at . Visitchat.deepseek.comand enableDeepThinkmode to interact with the full 671 - billion - parametric quantity poser .
Alternatively , you may reach the Zero theoretical account or any distilled variant via theHugging Face app , where you may download lightweight models to run topically on your computer .
Why DeepSeek R1 Matters
Outside of Microsoft ’s Phi 4 fashion model , there is n’t another open - source abstract thought model available . Phi 4 , however , has only 14 billion parameters and can not compete with OpenAI ’s o1 closed models . DeepSeek R1 allow a free , capable - seed alternative that equal closed - source option like o1 and Gemini 2.0 Flash Thinking . For developers , the price - effectiveness and open accessibility of R1 take a crap it especially appeal .
The only downside is that , as a Chinese - develop model , DeepSeek must comply with Chinese governance regulations . This means it wo n’t respond to sensitive topics like Tiananmen Square or Taiwan ’s independence , as the Cyberspace Administration of China ( CAC ) check that all responses align with “ core socialistic value . ”
Samsung Galaxy S25 FE Could Use This New Chipset (It’s…
iPhone To Ditch Flat Screen? New Leak Hints At Curved…
One UI 8 For Watch To Bring Now Bar And…
New Spotify Update Makes It Easier to Discover New Music:…
It’s Official: Samsung Galaxy S25 Edge Launch Date Out!
Samsung Galaxy Z Flip 7 FE Price Leaks But It…
10000mAh Battery?! Realme’s New Phone Aims to Replace Your Powerbank
Samsung Galaxy S26: Exynos Set to Make a Comeback
Goodbye Dynamic Island? Apple iPhone 18 Set to Introduce This…
Leaked! Samsung S25 Edge Price, Specs, Launch Date
