AI Potluck
Model components / Base / pretrained models

Kimi K3

Moonshot AI

Moonshot AI's flagship mixture of experts, announced 16 July 2026 with 2.8 trillion total parameters activating sixteen of 896 experts, native vision and a one-million-token context. It is built on Kimi Delta Attention and attention residuals, and runs on kimi.com, Kimi Work, Kimi Code and the Kimi API.

Consolidated 2026-07-29 from kimi-k2-6; openness follows K3, the governing release, whose weights landed 2026-07-27 ungated. Capability still describes K2.6, which is the release with a published benchmark table. Verified 2026-08-14 via the Kimi-K3 model card and its LICENSE file.

Openness

3 high confidence
3.0
weights
open(shipped 2026-07-27, ungated, 96 safetensors shards)
data
closed
code
partial(inference/deploy only
license
Modified-MIT(non-OSI

Modified-MIT weights - standard MIT below a 100M-MAU and $20M-monthly-revenue threshold - over a closed corpus, so 3. The Hugging Face API reports moonshotai/Kimi-K3 as neither gated nor private across 96 safetensors shards, so the weights are genuinely open and downloadable, and K3's own LICENSE file carries the Modified-MIT terms: MIT text, with the grant excepted for a licensee's commercial products or services having 'more than 100 million monthly active users, or more than 20 million US dollars' of monthly revenue, plus a prominent-display attribution clause. That is the same instrument K2.6 carries, read directly off K3 rather than assumed to continue from the K2.x line, which is why this reads with high confidence. Only inference and deployment code is published and no corpus is released, so the family lands at open weights rather than open source.

  • https://huggingface.co/api/models/moonshotai/Kimi-K3 recorded 2026-08-14

    `"gated":false`, `"private":false`, `license:other` in the repo tags with a LICENSE file in the sibling list, and 96 model-000NN-of-000096.safetensors shards, so the K3 weights shipped and download without a barrier. 1,871,575 downloads in the trailing 30 days; 10,621 likes. No corpus and no training pipeline appear in the tree - only configuration, processor and encoding modules.

  • https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE recorded 2026-08-14

    K3's own license file, read directly rather than assumed from K2.x. MIT text with the Modified-MIT carve-out - the grant is excepted where the licensee's commercial products or services have "more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies)" - and a prominent-display attribution requirement.

  • https://www.kimi.com/blog/kimi-k3 recorded 2026-07-19

    primary announcement — 2.8T MoE, 1M context, native vision; API live; full weights by 2026-07-27

  • https://huggingface.co/moonshotai/Kimi-K2.6 recorded 2026-07-19

    prior Kimi open-weight release under Modified MIT (the continuity the old record assumed from)

Adoption

4 high confidence
4.0

2,676,581 downloads in the trailing 30 days summed across the 2 declared artifact(s) (moonshotai/Kimi-K2.6 805,006; moonshotai/Kimi-K3 1,871,575). Bands at level 4 (1M-10M) on the software and model adoption scale. K3's weights shipped on 2026-07-27, so the band rests on a download count for the release that governs rather than on a production-traction reading.

Capability

5 high confidence
5.0

SWE-bench Verified 80.2%, SWE-bench Pro 58.6% (up from 50.7% in K2.5), Terminal-Bench 2.0 66.7% (up from 50.8%). Reported as tying GPT-5.5 on coding and beating some top US models, and specialized for long-horizon agentic work - 300-agent swarms and 4,000 coordinated steps. That is frontier-tier for agentic and coding tasks. The figures describe K2.6 rather than K3, the release that governs the family, because Moonshot has published no comparable benchmark table for K3; that is immaterial to the band, since 5 is the ceiling either way.

Verified 2026-08-13