
What happened
A homelabber detailed a second-gen Optane build with two added RTX 3090s that runs the 2.8-trillion-parameter Moonshot Kimi K3 model at 3.11 tokens per second, in a follow-up post on Reddit.
Why it matters
This shows that a very large model can be run on a home-built setup, though the persistent Optane memory still gets very hot and extra fans are needed, according to the homelabber.
Summaries like this, every morning.
This build is a follow-up to an earlier Optane-based AI setup, with the key change being the addition of two RTX 3090s. The homelabber shared the results in a Reddit post, showing that the 2.8-trillion-parameter Moonshot Kimi K3 model can run at 3.11 tokens per second on this home-built system. The use of a RAM drive to hold almost half the model is a notable tweak, yielding a 41% gain in generation speed and a 26% gain in prompt evaluation. However, the persistent Optane memory still runs very hot, so extra fans remain necessary, suggesting that cooling is an ongoing challenge for such dense setups.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Cisco Systems announced AI-focused alliances with DXC Technology and Qumulo on 19 November 2026

The Curious Analyst said Nvidia's leadership drama is already reflected in its share price, and that the compa…

Wells Fargo is adding Section 15 to its Online Access Agreement, effective with an updated agreement rolling o…

Visa expanded its Agentic Ready program — part of Visa Intelligent Commerce — to merchants, PSPs, acquirers an…

Marvell Technology and Credo Technology Group reached a confidential settlement resolving their patent dispute…

SMART-TD's National Health & Welfare Plan will lift the standard monthly contribution from $308.49 to $349.35…
