this post was submitted on 17 Jul 2026
33 points (82.4% liked)

Technology

86417 readers
3066 users here now

This is a most excellent place for technology news and articles.

Our Rules

  1. Follow the lemmy.world rules.
  2. Only tech related news or articles.
  3. Be excellent to each other!
  4. Mod approved content bots can post up to 10 articles per day.
  5. Threads asking for personal tech support may be deleted.
  6. Politics threads may be removed.
  7. No memes allowed as posts, OK to post as comments.
  8. Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
  9. Check for duplicates before posting, duplicates may be removed
  10. Accounts 7 days and younger will have their posts automatically removed.

Approved Bots

founded 3 years ago
MODERATORS
L3s@lemmy.world
enu@lemmy.world
technopagan@lemmy.world
L4s@lemmy.world
L3s@hackingne.ws
33
Chinese startup Moonshot AI unveils Kimi model it says rivals OpenAI, Anthropic (www.cnbc.com)
submitted 14 hours ago by themachinestops@lemmy.dbzer0.com to c/technology@lemmy.world
8 comments fedilink hide all child comments
 

Kimi K3 still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol on overall performance, the company said on Friday, but consistently outperformed other tested models.

The model beat Claude Opus 4.8 and GPT 5.5 — models that sit just behind Anthropic and OpenAI’s leading-edge systems — on benchmarks including coding and general agents, according to Moonshot.

It’s China’s largest AI model so far, with 2.8 trillion parameters, referring to the size of its neural network.

you are viewing a single comment's thread
view the rest of the comments
[–] chocrates@piefed.world 4 points 13 hours ago (3 children)

How big is un-quantized 2.8t param model?

I just got some hardware but I still don't have a ton of vram :(

[–] L_Acacia@lemmy.ml 8 points 13 hours ago* (last edited 13 hours ago)

Its trained at native MXFP4 with MXFP8 activations layer, so you need around 1.5TB of VRAM to fully offload it without taking into account the context cache. It might be doable to do some smart expert offloading and swapping, but expect minimum 500GB of VRAM and 1TB of system RAM minimum and the t/s would be reduced.

Its only realistically runnable on datacenter grade gpu at decent speed for now (and judging from the price of ram for the next few years)

[–] cecilkorik@lemmy.ca 3 points 12 hours ago

Hobbyists are working in the billions or tens of billions. Trillions is insane. I'm sure some people can run it, but not many.