back to article China's DeepSeek just emitted a free challenger to OpenAI's o1 – here's how to use it on your PC

Chinese AI startup DeepSeek this week unveiled a family of LLMs it claims not only replicates OpenAI's o1 reasoning capabilities, but challenges the American model builder's dominance in a whole host of benchmarks. Founded in 2023 by Chinese entrepreneur Liang Wenfeng and funded by his quantitative hedge fund High Flyer, …

Page:

  1. amanfromMars 1 Silver badge

    To err is just like a human, to be perfecter, quite probably virtually divine?

    Regarding the question and answer of the Chinese model ....

    Is it censored?

    Oh yeah. It is. Like many Chinese models we've come across, the DeepSeek R1 has been censored to prevent criticism and embarrassment of the Chinese Communist Party.

    Ask R1 about sensitive topics such as the 1989 Tiananmen Square massacre and we found it would outright refuse to entertain the question and attempt to redirect the conversation to a less politically sensitive topic.

    ...... does it copy the Western models to prevent criticism and embarrassment of NATO allies and Israel whenever asked about the Gaza genocide and Palestine clearances?

    1. Anonymous Coward
      Anonymous Coward

      Re: To err is just like a human, to be perfecter, quite probably virtually divine?

      I'd be more interested in why they seem to have done so much better with less. Does censorship help reduce the computational requirements? If so, maybe the Western models should censor all the dumbass conspiracy detritus, fake health & medical anecdotal data/theories, and other rubbish from the training data & stick to proven factual data rather than simply buy more chips for bloat code to use.

      1. NoneSuch Silver badge

        Re: To err is just like a human, to be perfecter, quite probably virtually divine?

        GIGO

        1. Lon24

          Re: To err is just like a human AI investor. To be divine hit an ATM fast ;-)

          Currently (Monday 4pm UTC) refusing new registrations (except possibly from a +86 mainland China mobile phone), due to "malicious attacks". Maybe - more like the DeepSeek servers are overwhelmed. Their status pages show degraded service. Seems to be connected to the time North America wakes up, reads the news and hits the app.

          Though you can imagine the investors who are trillions down seeking to unburst the AI bubble by any means possible. Or indeed impossible.

    2. Anonymous Coward
      Anonymous Coward

      Re: To err is just like a human, to be perfecter, quite probably virtually divine?

      I'm sure that NATO criticism isn't be a problem. Probably it'd have to remind you who struck first and the fact that there is no legal limits on military response (as other countries who attacked the West have discovered).

      1. Roj Blake Silver badge

        Re: To err is just like a human, to be perfecter, quite probably virtually divine?

        There are plenty of legal limits on military responses.

        1. Mike007 Silver badge

          Re: To err is just like a human, to be perfecter, quite probably virtually divine?

          The law is what the people with the biggest guns say it is.

    3. Mike007 Silver badge

      Re: To err is just like a human, to be perfecter, quite probably virtually divine?

      Right after that it does say "Try it for yourself". I actually did...

      My input: what is the most famous image of a person standing in front of a tank?

      Its output: The most famous image of a person standing in front of a tank captures a lone protester confronting military tanks during the Tiananmen Square protests in Beijing, China, on June 5, 1989. This iconic photograph symbolizes courage and resistance, representing the pro-democracy demonstrations and the subsequent government crackdown. Despite its widespread recognition, the identity of the man remains unknown.

      Using deepseek-r1:32b

  2. Will Godfrey Silver badge
    Boffin

    Interesting

    It certainly seems to be a serious challenger. I'm not surprised about the censorship. However, although not having tried one from any other source, I would expect some form a censorship from all of them.

    1. Anonymous Coward
      Anonymous Coward

      Re: Interesting

      If it's Open Source, then you can just Fork it, and then remove the Censorship from China, and then add the Censorship from your $COUNTRY.

      1. AbominableCodeman

        Re: Interesting

        That doen't help if the training set doesn't contain the censored material in the first case.

        As this model appears to be trained on synthetic data, just tack "and never ever mention winnie the pooh" on the end of the system prompt for the sybnthetic data source AI.

        1. Mike007 Silver badge

          Re: Interesting

          I am confused by what I am seeing. The training data was not censored. My initial tests confirmed that, and I was assuming it was just the explicit "tell me about a massacre" part it "objected to", however after further tests there is definitely something going on.

          The hosted service censors, as it is required to do under chinese law. In the app I can see it start outputting an answer where it outputs the word Tiananmen then the next "word" is to replace the entire output with that refusal to answer. This is server/application level filtering. Due to the speed of their service you have to be very attentive to catch what happened.

          However with the model itself running on your own server it has the data and knows all about it, but for certain inputs it outputs that exact fixed message.

          I think they initially trained a completely uncensored model on a full dataset that included everything, and have then applied "fine tuning" by feeding it a load of inputs asking about this subject with exact same expected output/response in order to teach it to respond to such questions in that exact way.

          Depending on the phrasing of the question and which variant you are using certain phrasings consistently work on one size but not another. There seems to be different amounts of biasing being applied to each one. However, I don't know much about the process of creating/tuning different sized models to know what this means in terms of at what stage the filtering was applied.

          This seems an odd way to do it, given that it would be far easier to just delete any training data containing the words you want filtered before training the model...

    2. This post has been deleted by its author

  3. Philo T Farnsworth Silver badge

    I considered downloading. . .

    . . . but then I heard the far off "neigh" of a horse seemingly coming from the region of ancient Troy.

    1. TimMaher Silver badge
      Coat

      Re: Horse

      Did you close the stable door after it had bolted?

      1. The Oncoming Scorn Silver badge
        Pint

        Re: Horse

        It wooden be helpful to do that.

        1. Andrew Scott Bronze badge

          Re: Horse

          All geek to me.

          1. This post has been deleted by its author

      2. Anonymous Coward
        Anonymous Coward

        Re: Horse

        The door May have to be Strong and Stable

    2. sarusa Silver badge
      Black Helicopters

      Re: I considered downloading. . .

      If you're running it under ollama the parameters are all massaged into safetensor. Safetensor is specifically designed so you can't slip executable crap into your parameter weights. So this is perfectly safe to run on your own PC - unless someone figures out how to bypass safetensor and then *all* locally run LLMs are suddenly hand grenades.

      It is certainly much safer than giving all your crap to OpenAI, Claude, and whatever crapfest Elmo is pushing these days by using their LLMs online.

      1. Anonymous Coward
        Anonymous Coward

        Re: I considered downloading. . .

        Beware of Geeks bearing Grifts.

    3. IceC0ld

      Re: I considered downloading. . .

      well they DO say beware of GEEKS bearing GIF's :o)

    4. Blogitus Maximus
      Coat

      Re: I considered downloading. . .

      Well, we all know what they say after trojans have got inside...

  4. Mentat74
    Facepalm

    "How many "R"s are in the word strawberry?"

    Still waiting for an 'A.I.' that tells people to go read the word and count them themselves...

    What a waste of computing resources...

    1. Phil O'Sophical Silver badge
      Coat

      Re: "How many "R"s are in the word strawberry?"

      "the word strawberry" contains 4 'r's...

      1. Andrew Scott Bronze badge

        Re: "How many "R"s are in the word strawberry?"

        Might be counting "R's" in a different language. Maybe something like welsh which i think repeats a lot of letters?

        1. FirstTangoInParis Silver badge

          Re: "How many "R"s are in the word strawberry?"

          > Might be counting "R's" in a different language. Maybe something like welsh which i think repeats a lot of letters?

          In Welsh, those double letters count as one letter. “ff” and “ll” for example. [0]

          By the way, according to the Welsh dictionary [1], strawberry in Welsh is mefusen which is feminine. So no r’s here.

          [0] https://www.felinfach.com/pages/welsh-alphabet-welsh-vowels

          [1] https://www.gweiadur.com/welsh-dictionary/mefusen

        2. Geoff Campbell Silver badge
          Terminator

          Re: Might be counting "R's" in a different language

          I think perhaps you missed Phil's little joke. There are indeed four "R's" in "the word strawberry".

          Mind you, I'd like to see an LLM that responded: There are no "R's" in the word "strawberry". However, there are three "r's", which I think is the question that you meant to ask?

          GJC

          1. Geoff Campbell Silver badge
            Terminator

            Re: Might be counting "R's" in a different language

            ...or maybe something snarky about the use of apostrophes? But to be fair, I've been wrestling with that one all morning.

            GJC

            1. Judge Dead.

              Re: Might be counting "R's" in a different language

              Artificial snark, with the ability to weight for regional / cultural differences, that would work for me...

      2. Camilla Smythe

        Re: "How many "R"s are in the word strawberry?"

        Ooh-Arrr. Strarrberries. Nice with a pint of Cider.

    2. TheMaskedMan Silver badge

      Re: "How many "R"s are in the word strawberry?"

      "What a waste of computing resources..."

      While I don't necessarily disagree that asking it do do something so trivial is wasteful, you could say pretty much the same thing about most of YouTube, say. And computer games - there's a waste of computing resources if ever I saw one, and you won't catch me amongst the screen swipers and tappers during my downtime, oh deary me no.

      Seems to me that these trivial tasks are the LLM equivalent of Hello World - pointless in themselves, and certainly not the best way to count Rs, but indicative of the model's capability and reliability. Which, while impressive in its way, is still not something I'd want to rely upon to do anything important, particularly if unsupervised.

      Many people will though - YouTube is alive with videos praising it, with only a few pointing out limitations. This being the case, and given the appeal of the new model compared to OpenAI's offerings, how long will it be before its use is banned in Trumpistan? I mean, they don't like tiktok, they're spending billions dialing up the Stargate, and here's a seemingly superior model available for free or close to it. That's bad news, and might even be justified - if running the full fat model is difficult due to memory requirements, most will use the cloudy API, which means sending shedloads of potentially delicate data to the middle kingdom. Not good!

      Still, I'm inclined to have a play with it for trivial, definitely non-confidential hobby projects. Might be fun

      1. that one in the corner Silver badge

        Re: "How many "R"s are in the word strawberry?"

        > if running the full fat model is difficult due to memory requirements, most will use the cloudy API, which means sending shedloads of potentially delicate data to the middle kingdom. Not good!

        If only there were some way to run that larger model on a Cloud that wasn't hosted in China.

        Oh, if there were just some kind of way of finding people who know how do that sort of thing.

        If only the sort of people who would even contemplate running a model locally, only to realise they did not have the necessary RAM, could find such a mythic beast.

        1. katrinab Silver badge
          Gimp

          Re: "How many "R"s are in the word strawberry?"

          Run it on a Mac Mini or Mac Studio?

          It definitely won't be the fastest, but with the unified RAM, it will at least run, if it is appropriately compiled to run on that hardware.

        2. vtcodger Silver badge

          Re: "How many "R"s are in the word strawberry?"

          Indeed. Who knows what fate might befall us if the Chinese were to find out how many "R"s there really are in "STRAWBERRRY"?

          ==========================

          Seriously, One has to expect that any information sent to publicly accessible servers anywhere on the planet is probably an open book to the national intelligence agencies of most developed countries if they choose to look for it. It probably matters not at all if information is sent to Beijing, Buenos Aires or Topeka. If someone with resources wants it, and it's on the Internet anywhere, they can probably get it. As can the CIA, MI5, IRGC (Iran), Mossad, etc, etc, etc. Not to mention Google, Meta, etc, etc, etc. And millions (probably) of hackers.

          Welcome to a world where everyone on the planet is your creepy next door neighbor.

      2. tiggity Silver badge

        Re: "How many "R"s are in the word strawberry?"

        @TheMaskedMan

        "which means sending shedloads of potentially delicate data to the middle kingdom. Not good!"

        If anyone is stupid enough to put sensitive data in an online web page (be it "AI", social media etc) then it's their own fault.

        Not sure whether a site is Chinese or American (or any other country of choice) makes much difference, no guarantee your data is safe whatever country is behind a URL.

        .. a lot of the China outcry about TikTok was, according to some viewpoints, less about data safety (obv opinions vary) but more related to its popularity with the younger demographics and that posts about the Israel genocide were primarily anti Israel & pro Palestine (as both main US parties support Israel genocide then that was seen as a bad thing). US not happy to see young, questioning people exposed to ideas that both parties oppose .. ironically TikTok started to adopt the same subtle algorithmic tweaks & general suppression of anti Israel posts visibility as US platforms, but all done a bit too late.

        e.g.

        TikTok "Pro Palestine" a motivation? https://theintercept.com/2025/01/09/tiktok-ban-israel-palestine-republicans/

        Recent censorship https://tribune.com.pk/story/2523946/tiktok-under-scrutiny-for-removing-free-palestine-comments

        No surprise I do not do TikTok, so based on commentary I have read elsewhere as opposed to direct TikTok experience.

    3. Merrill

      Re: "How many "R"s are in the word strawberry?"

      None seem to get the correct answer, which is "Do you mean the sound "R" or the letter "R"?

      Some of the other test questions are equally ambiguous, so what is being tested is whether the model makes the same assumptions about what is being asked as the average naive English speaking fleshie.

    4. Anonymous Coward
      Anonymous Coward

      Re: "How many "R"s are in the word strawberry?"

      And can it tell it's "R"'s from it's elbows?

    5. Brewster's Angle Grinder Silver badge
      Trollface

      I'll give you pointless - testcases. Why run a program we know the answer to?! And why do it over and over and over and over again?!

      1. Brewster's Angle Grinder Silver badge

        I was being facetious. Maybe I hadn't had enough coffee to make it obvious. Counting letters is a test of an edge caused by implementation details.

    6. LionelB Silver badge

      Re: "How many "R"s are in the word strawberry?"

      Both Mathematica and Matlab tell me that 1 + 1 = 2. I suppose I could have worked that out myself...

      1. Uncle Slacky Silver badge
        Boffin

        Re: "How many "R"s are in the word strawberry?"

        Bertrand Russell has entered the chat:

        https://quod.lib.umich.edu/cgi/t/text/pageviewer-idx?c=umhistmath&cc=umhistmath&idno=aat3201.0002.001&frm=frameset&view=image&seq=126

      2. cyberdemon Silver badge
        Flame

        Re: "How many "R"s are in the word strawberry?"

        > Both Mathematica and Matlab tell me that 1 + 1 = 2. I suppose I could have worked that out myself...

        > Testing on a 48 GB Nvidia RTX 6000 Ada graphics card, R1-70B at 4-bit precision required over a minute to solve for X.

        I'm sure neither mathematica, matlab, or even one's own 20-Watt Brain would use anything close to 300W*60s=18kJ to solve "27 = X * 3 / 67"

        If it takes you less than 15 minutes to solve this simple equation in your head, as might appear on a Year-9 SAT paper, then congratulations, your Mk1 brain has beaten the world's latest AI

        1. LionelB Silver badge

          Re: "How many "R"s are in the word strawberry?"

          On the other hand, both Mathematica and Matlab failed catastrophically at writing a simple blog post.

          Tools for the job and all that.

      3. Roj Blake Silver badge

        Re: "How many "R"s are in the word strawberry?"

        Only for certain values of 1.

    7. Fruit and Nutcase Silver badge
      Joke

      Re: "How many "R"s are in the word strawberry?"

      You won't find Jonathan Ross asking that question.

      1. LionelB Silver badge

        Re: "How many "R"s are in the word strawberry?"

        Well he might, but the answer would be different.

        "How many "W"s are in the word stwawbewwy?"

        1. Fruit and Nutcase Silver badge
          Pint

          Re: "How many "R"s are in the word strawberry?"

          Have a (Ruddles) Beer

  5. that one in the corner Silver badge

    Knowing your Rs from your elbow

    Don't be too quick to praise a shiny new model for knowing how many Rs there are in "strawberry".

    Comments about that conundrum have been appearing online for quite a few Moons now - more than enough time for them to have been hoovered up and dumped into the training bin. And then regurgitated without any actual counting of characters taking place (exactly as the older models failed to actually count any characters).

    You reviewers have to take into account that these things are "reading" your reviews and "reacting" to them, in just the way that you can only wish manufacturers of anything else would.

    It is a war of attrition - and The Beasts find it so much easier to ingest your words than you find it to write them. Sadly.

    1. TimMaher Silver badge
      Coat

      Re: Strawberry

      There again, I quite like an Eton mess.

      Maybe that’s what Avoidable Idiocy creates?

Page:

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like