Can you own it?

Intended & out-of-scope use

Kimi is Moonshot AI's model line; the confirmed open release is Kimi K2, a trillion-parameter Mixture-of-Experts model (1T total / 32B active) built with an explicit agentic, tool-use and coding focus. Its intended use is frontier-adjacent assistant, agentic, and coding workloads where you want an open-weight model rather than a closed API - and where you can afford the serving infrastructure.

For adoption, deploy the Kimi-K2-Instruct (safety-tuned) variant; Kimi-K2-Base is a foundation checkpoint for research and fine-tuning and is out of scope for customer-facing use. Because it is a China-based release with China-aligned topic censorship, no companion guard model, and no EU AI Act documentation package, high-stakes, autonomous, or EU-regulated decision-making is out of scope without the external control stack described in the Implement domain and a self-assembled compliance package.

Known limitations, bias & failure modes

• China-aligned topic censorship. Like other China-based open-weight models, Kimi K2 applies content filtering aligned with Chinese content rules on politically sensitive topics
  • a behavioural quirk to account for, especially for non-coding prompts.
• Lighter safety tuning, no guard model. Instruct alignment withstands casual jailbreaks but coverage is lighter than Western frontier labs and there is no companion guard/classifier.
• Trillion-parameter operational burden. There are no small variants; the smallest viable deployment is a multi-node cluster even in block-FP8 (~1 TB of weights).
• Base checkpoint is untuned. Kimi-K2-Base has no safety tuning and should be treated as a research artifact.

The offsetting advantage is capability: on independent agentic/coding benchmarks Kimi K2 is among the strongest open-weight models available.

Openness tier & components

Kimi K2 is open_weights, not open-science. Weights (Base and Instruct) are downloadable under a permissive modified MIT license with a detailed model card and technical report, so weights, documentation and license are Open. But the 15.5T-token training data and the training code are not released, and evaluation is only partially reproducible - so those components are Closed/Partial. You can run, adapt and redistribute the model, but you cannot reproduce it.

License terms & permitted use

Kimi K2 ships under a Modified MIT License: standard MIT permissions (use, modify, redistribute, commercialize) plus one attribution clause - if the software or a derivative is used in a commercial product or service with more than 100 million monthly active users or more than $20 million USD in monthly revenue, you must prominently display "Kimi K2" on that product's user interface. For nearly all adopters the license is effectively permissive MIT; only large-scale deployers trigger the attribution requirement. Two caveats: the modification means it is not OSI-certified, and you should verify the LICENSE file per checkpoint. This clause is a load-bearing input to the legal (3) score.

Supply-chain & provenance

Weights are distributed from the verified moonshotai org on Hugging Face as block-FP8 safetensors (no pickle requirement) with per-file checksums. The checkpoint trust checklist scores 5/8; the missing controls include cryptographic weight signing (Sigstore/model-signing) and SLSA build attestation, which is why provenance is a 4, not a 5. No incidents or malicious-mirror findings are on record for the canonical org. Because the weights are large and distributed in block-FP8, verify per-file checksums and pin the exact revision on download.

EU AI Act posture

Kimi K2 is a GPAI model. On the active-parameter compute estimate (~3×10²⁴ FLOPs: 32B active × 15.5T tokens) it sits under the 10²⁵-FLOPs systemic-risk threshold - so, unlike the largest systemic-risk peers, there is no Article 55 hard flag and it plausibly reaches the Article 53 open-source exemption from the Annex XI/XII technical-documentation duties. Two caveats keep this at partial: the modified MIT license is not an OSI-certified free/open license, and the surviving obligations - a copyright policy and a public training-content summary - are not published, and cannot be reconstructed because the corpus is closed. A China-based provider is unlikely to furnish an EU documentation package, so an EU deployer must self-assemble compliance material and inherits these gaps if it places a derivative on the market.

Benchmarks & evaluation

Kimi K2 is among the strongest open-weight models on agentic and coding benchmarks (for example SWE-bench Verified ≈ 65.8, plus strong tool-use/agentic scores), frontier-adjacent rather than merely in-class. OneHill did not run its own benchmarks this session; the figures are aggregated from Moonshot's technical report and independent evaluation (ev-perf, ev-tech-report). This is marked partial because the results are third-party rather than OneHill-reproduced and the evaluation harness is only partially open.