Tier 3 Practical open-weight
What the publisher claims: open-source, open-weight
The one-liner: Kimi Delta Attention model released under plain MIT, unlike the modified MIT of Kimi K2.
The six questions
weights downloadable
MIT (differs from the modified MIT attached to Kimi K2)
runs with transformers trust-remote-code and vLLM
architecture, token counts and evals published; corpus withheld
Limited to moderate.
The MIT-licensed Kimi-Linear weights and the KDA kernel in flash-linear-attention sit on public GitHub where outside PRs are sometimes merged, but the model roadmap is set by Moonshot.
License
MIT (HF license:mit tag on Kimi-Linear-48B-A3B-Base and Instruct) OSI-approved — license source ↗
What is actually released
- base and instruct weights (48B total, 3B active, 1M context)
- KDA kernel open-sourced in flash-linear-attention
- tech report arXiv 2510.26692
- evaluation data
Why it matters
A useful within-family licence contrast for the same vendor, showing that branding the whole vendor open-weight would have been wrong.
Sources
- https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct
- https://huggingface.co/api/models/moonshotai/Kimi-Linear-48B-A3B-Base
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.