Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepSeek-V3

github.com

Post reply on HN
Print view View original

21–30 of 44 posts

Re: DeepSeek-V3

#21

Post by cloverich »

chvid wrote: post #8
msoad wrote: post #3

You can run a model that can beat 4o which was released less than 6 months ago _locally_! I know this requires a ton of hardware but OpenAI will not be the leader in 2025 I can assume. Always bet on open source (or rather somewhat more open development strategies) The math and coding performance is what we really care about. I am paying for o1 Pro and also Sonnet, in my experience beside Sonnet being faster, it is al…

OpenAi toppled as LLM leader by an open source / open weight company? OpenAi has much more capital and compute than any of its competitors (especially deepseek); if that was to happen it would demonstrate that capital and compute doesn't matter as much as it is assumed ... (and it just might be the thing that pops the current ai bubble).

Until the models can host themselves theyll always need a company to make the experience good enough for typical users; OpenAI can always host open source models instead of their own and their user base will mostly stick around, especially if they can leverage their existing base into a network effect. I wouldnt be surprised it they are investing heavily into this vs pure model hosting running.

Im thinking their real challenge will be surviving Apple (once they go all in) or Google (if they can figure out how to make a good product). Or something along those lines.

Re: DeepSeek-V3

#23

Post by xur17 »

rubslopes wrote: post #20

Already available at OpenRouter: https://openrouter.ai/deepseek/deepseek-chat Cost / million tokens: Input $0.14 Output $0.28

For comparison, Claude 3.5 Sonnet (my favorite model for coding tasks) is: Input $3 Output $15.

Re: DeepSeek-V3

#24

Post by 383toast »

apexalpha wrote: post #19
polotics wrote: post #9

Earlier quoted context omitted.

well yes, locally, if you assume that someone's got about 300'000 dollars of hardware at hand... right? as you are not paying for Gemini, may I ask why, did you try it and find it inferior?

I bought two (relatively) old datacenter GPUs with 48gb VRAM total for €200 that gets me 7 token/s for a 70b model.

which GPUs?

Re: DeepSeek-V3

#25

Post by rahimnathwani »

Pricing per million tokens:

  Model               Input     Output    
  ─────────────────────────────────────
  Claude 3.5 Sonnet   $3.00     $15.00
  GPT-4o              $2.50     $10.00
  Gemini 1.5 Pro      $1.25      $5.00
  Deepseek V3         $0.27      $1.10
  GPT-4o-mini         $0.15      $0.60

Re: DeepSeek-V3

#27

Post by boroboro4 »

janice1999 wrote: post #18

> a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. What kind of hardware do you need to run this?

They discuss it in the paper and recommend 32 GPUs (H800 in their case) for prefill stage and 320 GPUs for decoding.

=)

Re: DeepSeek-V3

#28

Post by zargon »

383toast wrote: post #24
apexalpha wrote: post #19

Earlier quoted context omitted.

I bought two (relatively) old datacenter GPUs with 48gb VRAM total for €200 that gets me 7 token/s for a 70b model.

which GPUs?

Not the GP, but I bought a few P40s over the summer for $150 each. Last I checked they're more expensive now, but it's still cheap vram and fast enough at inference for me.

Re: DeepSeek-V3

#30

Post by miletus »

handzhiev wrote: post #26

How is this not on the front page. It's a remarkable release.

Noticed the same thing. DeepSeek-V3 is remarkable (beats 4o/ claude), but it's not on the front page.

It seems they don't want china to win haha

Post reply on HN