Kimi K3

weights

open(shipped 2026-07-27, ungated, 96 safetensors shards)

data

closed

code

partial(inference/deploy only

license

Kimi-K3-License(non-OSI

Kimi K3's weights are ungated on Hugging Face under the Kimi K3 License, an MIT-style grant with two conditions. A company running a model-as-a-service business with more than $20 million in revenue over any 12 months must sign a separate agreement with Moonshot AI before commercial use, and a product with more than 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3". Only inference and deployment code is published and no training corpus is released, so the family counts as open weights rather than open source.

• https://huggingface.co/api/models/moonshotai/Kimi-K3 recorded 2026-10-11

  `"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. No corpus and no training pipeline appear in the tree - only configuration, processor and encoding modules.

• 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)

• https://huggingface.co/moonshotai/Kimi-K3/raw/main/LICENSE recorded 2026-10-11

  Titled "Kimi K3 License". Section 2: a licensee that "operates a Model as a Service business", with aggregate revenue that "exceeds 20 million US dollars" over "any consecutive 12 months", "must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose". Section 3: a commercial product with "more than 100 million monthly active users" or "more than 20 million US dollars" in monthly revenue must display "Kimi K3". Section 4 exempts internal use. The grant itself carries no user or revenue exception.

• 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

Adoption is measured as Hugging Face downloads summed across the two declared artifacts, Kimi-K2.6 and Kimi-K3. K3's weights shipped on 2026-07-27, so the reading reflects downloads for the release that now governs the family rather than a longer production-traction record.

• https://huggingface.co/api/models/moonshotai/Kimi-K2.6 recorded 2026-08-13

  805,006 downloads in the trailing 30 days for moonshotai/Kimi-K2.6

• https://huggingface.co/api/models/moonshotai/Kimi-K3 recorded 2026-08-13

  1,871,575 downloads in the trailing 30 days for moonshotai/Kimi-K3

Kimi K3 scores 88.3 on Terminal-Bench 2.1, 67.5 on DeepSWE and 93.5 on GPQA Diamond in Moonshot's published evaluation, within a few points of Claude Fable 5 and GPT-5.6 Sol on most rows of the same table and ahead of them on some. The coding results were run with Moonshot's own Kimi Code harness rather than an independent one.

• https://benchlm.ai/models/kimi-2-6 recorded 2026-09-29

  Third-party page for Kimi K2.6 still listing "Software Engineering Benchmark Verified" at 80.2 and SWE-bench Pro at 58.6, both labelled "Provider exact".

• https://huggingface.co/moonshotai/Kimi-K3/raw/main/README.md recorded 2026-09-29

  K3 model card evaluation table, a K3 column against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and GLM-5.2, with the note that on DeepSWE "under which Kimi K3 attains 67.3 with the mini-SWE-agent harness" and that for Terminal-Bench 2.1 "Kimi K3 is evaluated with the Kimi Code harness". K3 rows read GPQA Diamond 93.5, DeepSWE 67.5, Terminal-Bench 2.1 88.3, BrowseComp 91.2 and OSWorld-Verified 84.8.

• https://officechai.com/ai/kimi-k2-6-benchmarks/ recorded 2026-09-29

  Press write-up of the K2.6 release - "K2.6 edges out or closely matches GPT-5.4 (xhigh) and Claude Opus 4.6 (max effort)" - naming SWE-Bench Pro 58.6 as the lead result.

• https://www.kimi.com/blog/kimi-k2-6 recorded 2026-09-29

  Moonshot K2.6 post (redirects to kimi.ai), titled "Kimi K2.6 Tech Blog: Advancing Open-Source Coding", with a benchmark sheet listing K2.6 at SWE-Bench Verified 80.2, SWE-Bench Pro 58.6 and Terminal-Bench 2.0 66.7, and noting "Terminal-Bench 2.0 scores were obtained with the default agent framework (Terminus-2)".

• https://www.kimi.com/blog/kimi-k3 recorded 2026-09-29

  K3 announcement carrying the same evaluation table and notes, including "Kimi K3 is evaluated with the Kimi Code harness. For all other models, we report the best score across harnesses".