SURVEY RECORD N° 14 FILED 2026-08-13
← RETURN TO THE REGISTER

Kimi K3

Largest open release to date (~2.8T MoE); the current open frontier-substitute pick.

Context~1M
Passage, per 1M in / out~$0.6 / ~$3
AA Intelligencenot yet surveyed
LicenseModified MIT
ModalitiesText, image
Official sitemoonshot.ai →

In its favor

  • Among the largest open releases (~2.8T MoE)
  • Adds native vision
  • ~1M-token context window
  • Current open frontier-substitute pick

Against it

  • Benchmark figures are vendor-reported, not independent
  • Modified MIT license — not fully permissive
  • Hosted price varies by provider
  • Very large to self-host at full size
open-weightvision1M contextreasoningvendor-benchmarks

ASSESSMENT

Kimi K3 is among the largest open releases to date — a mixture-of-experts around 2.8 trillion parameters — and it adds native vision to the open-weight frontier, with a roughly 1M-token context. For teams seeking an open substitute near the closed frontier, it is one of the current picks.

Two caveats shape the read. Its benchmark figures are vendor-reported rather than independently verified, so we hold its exact standing as provisional. And the license is a modified MIT rather than fully permissive — worth reading closely before commercial deployment. Hosted pricing varies by provider, and the model's sheer size makes full self-hosting demanding.

For organizations that want open weights with vision and are comfortable validating capability themselves, K3 is a serious option; those needing independently verified performance may prefer DeepSeek V4-Pro until third-party numbers land.

FIELD NOTE — Adds native vision. Hosted price varies by provider.

VISIT OFFICIAL SITE →