Tier 2 Restricted open-weight
What the publisher claims: open-weight, open source
The one-liner: Frontier-scale open-weight family with custom or modified terms and incomplete training transparency.
The six questions
The full K3 checkpoint and the K2-family releases are downloadable from Hugging Face and GitHub, with the 47-page technical report published alongside the weights.
K2-family releases use a Modified MIT with attribution and display conditions, and K3 replaces it with the custom Kimi K3 License, which requires a separate agreement for Model-as-a-Service businesses above $20M aggregate revenue and a prominent model-name display above 100M MAU or $20M monthly revenue; internal use is exempt but the conditions are real.
Self-hostable with published vLLM, SGLang, and TokenSpeed recipes; at roughly 1.5TB of weights K3 targets data-centre clusters, though reports of consumer RTX 5090 clusters running it exist.
The K3 technical report runs 47 pages across architecture (Kimi Delta Attention, Attention Residuals, Stable LatentMoE), pre-training, post-training, infrastructure, evaluations, and case studies, and includes third-party leaderboard results, but training-data composition is not published.
Limited: the report describes the method in enough detail to study but not to rebuild the model from published data.
Moonshot publishes Kimi weights, deployment recipes, and a 47-page technical report on Hugging Face with no open development process beyond filing issues on its own repositories.
License
Modified MIT for the K2 family; custom Kimi K3 License for Kimi K3 — license source ↗
- K3 requires a separate agreement with Moonshot for Model-as-a-Service businesses above $20M aggregate revenue over any 12 months
- K3 requires the model name displayed prominently above 100M MAU or $20M monthly revenue
- K2-family Modified MIT adds attribution and display conditions and a synthetic-data clause
What is actually released
- weights (K2, K2.5, K2.6, and the 2.8T-parameter K3 in MXFP4)
- 47-page technical report (K3)
- deployment recipes for vLLM, SGLang, and TokenSpeed
- third-party evaluation results
Why it matters
Kimi combines frontier-scale capability and practical self-hosting with a license and transparency profile that is not the same as fully open source, and its 2026 scale makes the MaaS condition bite far more companies than it would for a small model.
Sources
- https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE
- https://venturebeat.com/technology/kimi-k3s-full-weights-are-here-but-theyre-open-with-a-caveat-what-enterprises-should-know
- https://zentor.ai/blog/kimi-k3-technical-report
- https://huggingface.co/moonshotai/Kimi-K2.5
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.