Moonshot’s Kimi K3 ships open weights, but self-hosting needs 1.4TB and 18+ GPUs
Common_Dream9420 · reddit · 2026-07-27
- Moonshot released Kimi K3 open weights, described here as the largest open-weight model yet.
- The post says it has 2.8T parameters, a Modified MIT license, 91.2% on BrowseComp, and the best published agentic score at release.
- It also argues that despite the “open” label, self-hosting is impractical for most teams: the weights reportedly need 1.4 TB of storage and 18+ enterprise GPUs just to load.
- The author’s takeaway is that most people will use the API, so “open weights” still do not equal control over inference or data.
Related event: Moonshot releases Kimi K3 open weights amid license debate(155 posts)→
More from Infra
- 12 KV Cache Reduction Techniques Every AI Engineer Should Understand, Explained — blaizedsouza · 2026-09-11
- The shadow GPU capacity market is formalizing, with Meta selling excess compute to outside buyers — DavidLinthicum · 2026-09-11
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11