Kimi K3 is downloadable. That doesn’t mean you can run it.

Source: Nate’s Newsletter Date: 2026-07-20 URL: https://natesnewsletter.substack.com/p/kimi-k3-open-weights-cost

Summary

Kimi K3’s open weights (~2.8T total params / ~50B active, 16-of-896 experts) aren’t downloadable yet — Moonshot’s promised publish date is 2026-07-27. Nate’s core argument: even once they land, “open” doesn’t mean cheap to run. Moonshot’s own deployment guide recommends at least 64 high-end AI chips plus specialized cooling, networking, and power to serve the model, so K3 preserves demand for expensive hardware even as it erodes pricing power for closed-model vendors like OpenAI and Anthropic.

Implications

Feeds the open ≠ local / open decoupling from yours-to-run thread this radar has been tracking (GLM-5.2 753B → Inkling 975B → K3 2.8T — active params stay efficient but total params keep climbing into cloud-only territory). K3 is the sharpest data point yet: even the vendor shipping the open weights says you need a 64-chip cluster to run them. For local-first practice, K3 lands as a derivative/quant story at best, not a self-hostable model — same bucket as GLM-5.2, nowhere near reference 36GB-class hardware. Watch whether the 07-27 date holds (it’s a lab promise, not yet a shipped artifact).