DeepSeek's distilled new R1 AI model can run on a single GPU | TechCrunch

Comments

• LodeMike@lemmy.today ⁨2⁩ ⁨months⁩ ago

  So can a lot of other models.

  “This load can be towed by a single vehicle”

  source
• blarth@thelemmy.club ⁨2⁩ ⁨months⁩ ago

  7b trash model?

  source

  • vhstape@lemmy.sdf.org ⁨2⁩ ⁨months⁩ ago

    the Chinese AI lab also released a smaller, “distilled” version of its new R1, DeepSeek-R1-0528-Qwen3-8B, that DeepSeek claims beats comparably sized models on certain benchmarks

    Most models come in 1B, 7-8B, 12-14B, and 27+B parameter variants. According to the docs, they benchmarked the 8B model using an NVIDIA H20 (96 GB VRAM) and got between 144-1198 tokens/sec. Most consumer GPUs probably aren’t going to be able to keep up with

    source

    • avidamoeba@lemmy.ca ⁨2⁩ ⁨months⁩ ago

      It proved sqrt(2) irrational with 40tps on a 3090 here. The 32b R1 did it with 32tps but it thought a lot longer.

      source

      • -> View More Comments
    • brucethemoose@lemmy.world ⁨2⁩ ⁨months⁩ ago

      Depends on the quantization.

      7B is small enough to run it in FP8 or a Marlin quant with SGLang/VLLM/TensorRT, so you can probably get very close to the H20 on a 3090 or 4090.

      source
  • knighthawk0811@lemmy.world ⁨2⁩ ⁨months⁩ ago

    it’s distilled so it’s going to be smaller than any non distilled of the same quality

    source
  • LainTrain@lemmy.dbzer0.com ⁨2⁩ ⁨months⁩ ago

    I’m genuinely curious what you do that a 7b model is “trash” to you? Like yeah sure a gippity now tends to beat out a mistral 7b but I’m pretty happy with my mistral most of the time if I ever even need ai at all.

    source
  • TropicalDingdong@lemmy.world ⁨2⁩ ⁨months⁩ ago

    Yeah idk. I did some work with deepseek early on. I wasn’t impressed.

    HOWEVER…

    Some other things they’ve developed like deepsite, holy shit impressive.

    source

    • double_quack@lemm.ee ⁨2⁩ ⁨months⁩ ago

      Save me the search, please. What’s deepsite?

      source

      • -> View More Comments
• fogetaboutit@programming.dev ⁨2⁩ ⁨months⁩ ago

  ew probably still censored.

  source

  • T156@lemmy.world ⁨2⁩ ⁨months⁩ ago

    The censorship only exists on the version they host, which is fair enough. If they’re running it themselves in China, they can’t just break the law.

    If you run it yourself, the censorship isn’t there.

    source

    • jaschen@lemm.ee ⁨2⁩ ⁨months⁩ ago

      Untrue, I downloaded the vanilla version and it’s hardcoded in.

      source
    • MonkderVierte@lemmy.ml ⁨2⁩ ⁨months⁩ ago

      Yeah, i think the censoring in the LLM data itself would be pretty vulnerable.

      source
  • Mwa@lemm.ee ⁨2⁩ ⁨months⁩ ago

    You can self host it right??

    source

    • jaschen@lemm.ee ⁨2⁩ ⁨months⁩ ago

      The self hosted model has hard coded censored content.

      source
    • fogetaboutit@programming.dev ⁨2⁩ ⁨months⁩ ago

      if the model is censored… then what, retraining it? Or doing it from scratch like what open-r1 is doing?

      source