Kimi K3
the strongest open-weight model shipped, under the most conditional licence on this page.
it debuted at number one on lmarena's frontend code arena and it is not open source — third place is where those two facts meet.
the weights landed on 27 july: 2.8 trillion total parameters with 104 billion active, routing 16 of 896 experts through a stable latentmoe architecture, a one-million-token context, and genuinely native multimodality — text, image and video through a moonvit-v2 encoder rather than a bolted-on adapter. it shipped with a technical report and three infrastructure releases, and passed 99,000 downloads in its first day.
the capability is not in doubt. it debuted at number one overall on lmarena's frontend code arena, a seventeen-place jump from k2.6, beating claude fable 5 in six of seven frontend domains, with 67.5 on deepswe and 91.2 on browsecomp. moonshot's page claims first of 98 among open-weight models above 150b on the artificial analysis intelligence index at a score of 57; press coverage of that same index disagrees with itself about the overall position, placing it third in a headline and fourth in the body, behind fable 5 and gpt-5.6 sol.
the description above is ours, condensed from the ranking. pricing moves — check it on the vendor's own page before you rely on it.
- category
- open-weight llms
- pricing
- free (kimi k3 license)
- website
- huggingface.co
it debuted at number one on lmarena's frontend code arena and it is not open source — third place is where those two facts meet.
we researched this category against vendors' own pricing pages and licence files. that is where this line comes from — not from the vendor, and not from anything they paid for.