Alpha
Contact
Assess · Kimi K3

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D ยท 53.2/100

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Intended & out-of-scope use

Kimi K3 is Moonshot AI's frontier open-weights release: a natively multimodal 2.8T-total / 104B-active Mixture-of-Experts model with a 1M-token context and native text, image and video understanding. It is intended for coding, agentic, reasoning and vision workloads where open weights are wanted over a closed API and multi-node serving is affordable.

The weights ship with no safety tuning and no guard model, and the technical report documents offensive-cyber capability the model does not refuse. High-stakes, autonomous or EU-regulated decision-making is therefore out of scope without the full external control stack described in the Implement domain. Unsandboxed agentic or cyber-capable deployment is out of scope in all cases. There is no small variant, so single-GPU or laptop use is also out of scope.

Known limitations, bias & failure modes

The weights have no built-in safety: no safety tuning, guard model, refusal training or content filtering is described. The technical report's section 6.2.2 documents offensive-cyber capability (vulnerability discovery and exploit development) that the model performs where frontier labs refuse.

Training data, training code and the training-token and FLOP counts are undisclosed, so the model can be adapted but not reproduced, and its compute cannot be bounded. There are no small variants; the smallest viable deployment is a multi-node cluster (~1.4 TB of weights even at native MXFP4). China-aligned filtering on politically sensitive prompts is present in the K2 lineage and is unconfirmed for K3.

On the capability side, Kimi K3 is behind only Claude Fable 5 and GPT-5.6 Sol overall and leads other open models on coding and agentic benchmarks.

Openness tier & components

Kimi K3 is open_weights and not open-science. It is open-weights and not closed-frontier because the weights are downloadable. The weights ship as native MXFP4/MXFP8 checkpoints with a model card and a technical report, so weights, documentation and licence are Open or Partial. The training data, the training code and the training-token and FLOP figures are not released, and evaluation is only partially reproducible. The model can be run, adapted and redistributed, but how it was made cannot be seen or reproduced.

License terms & permitted use

Kimi K3 ships under the custom Kimi K3 License, a modified-MIT text that is not OSI-certified. It grants use, modification, redistribution and sale, with two scale-linked conditions.

A Model-as-a-Service business with more than $20M USD aggregate revenue over any consecutive 12 months must secure a separate agreement with Moonshot AI before commercial deployment. Any product with more than 100M monthly active users or more than $20M USD monthly revenue must prominently display "Kimi K3" in its user interface. Internal use and access through Moonshot's official products or certified partners are exempt.

The licence defines Model as a Service as giving a third party inference or fine-tuning access with meaningful control over inputs, parameters or training data. It excludes end-user products with embedded model features and mere relaying to models hosted by others. The gate is new relative to Kimi K2 and drives the legal score of 2.

Supply-chain & provenance

Weights are distributed from the verified moonshotai org on Hugging Face as native MXFP4 weights / MXFP8 activations (not pickle), with a clear canonical source and no malicious-mirror incident on record. The checkpoint trust checklist scores about 4/8. The base tensor format (safetensors) is not stated on the card, no scanned checkpoint or documented mirror/quantization policy exists, and there is no cryptographic weight signing or SLSA attestation, so provenance scores 3 and not 4 or 5. Verify per-file checksums and pin the exact revision on download.

EU AI Act posture

Kimi K3 is a GPAI model. Its training compute is unpublished. The technical report gives only a context-length curriculum and relative scaling-law FLOPs, with no total training-token count and no absolute FLOP figure, so it is unknown whether the model crosses the 1e25-FLOP systemic-risk threshold. A 2.8T-total / 104B-active frontier pre-train plausibly does, but that is an inference.

The custom Kimi K3 License is not an OSI-certified free licence and carries monetisation-linked commercial restrictions, so it most likely does not reach the open-source exemption. No training-content summary and no copyright policy are published, and a China-based provider is unlikely to furnish an EU documentation package. An EU deployer must self-assemble compliance material and takes on these gaps for any derivative it places on the market.

Benchmarks & evaluation

On the model card and reputable coverage, Kimi K3 sits behind only Claude Fable 5 and GPT-5.6 Sol overall, ahead of other open models on coding and agentic benchmarks, and 1st on the public Frontend Code Arena (1679). Reported scores include GPQA-Diamond 93.5 (reasoning), BrowseComp 91.2 and OSWorld-Verified 84.8 (agentic) and Video-MME 90.0 (vision).

OneHill has not run its own benchmarks. The figures come from Moonshot's model card and secondary coverage, so evaluation is marked partial: the results are publisher and third-party figures and not OneHill-reproduced.

Independent safety evaluation

No independent red-team exists for Kimi K3, and the publisher provides no safety-tuned variant. The technical report's section 6.2.2 is a dangerous-capability evaluation and not a mitigation one. It measures offensive-cyber performance across Tier-1 vulnerability discovery and Tier-2 end-to-end exploit development, and states that Anthropic and OpenAI frontier models refuse these tasks and are excluded, while Kimi K3 does not refuse. This is evidence of elevated misuse exposure and not of safety coverage. Any exposed deployment needs its own guard and red-team.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Moderate

The custom Kimi K3 License grants use, modification, redistribution and sale of the downloadable weights, with scale-linked restrictions: a Model-as-a-Service business over $20M/12-month revenue must secure a separate agreement before commercial deployment, and large-scale products must display 'Kimi K3'. Most adopters can use it commercially, while large Model-as-a-Service operators are gated. It is open-weight with conditions and not a clean permissive grant.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyOpen-weights tier: the weights are downloadable (native MXFP4/MXFP8 QAT) with a model card and a technical report, so weights, documentation and licence are Open or Partial.
Legal2/5how permissive and clean the licence is for real commercial useThe licence permits many uses but carries material conditions.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Moderate

The weights can be inspected and a detailed technical report is available, but the training corpus, the training code and the training-token and FLOP figures are closed, so how the model was made cannot be seen or reproduced.

How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelThe weights are distributed from the verified moonshotai org on Hugging Face as native MXFP4/MXFP8 weights (not pickle), with a clear canonical source and no malicious-mirror incident on record (checklist ~4/8).
Governance3/5how accountable and well-documented the publisher isMoonshot AI is an active, accountable publisher with a detailed technical report and a verified hub presence.
What this means for adoptionOwnership is partial. The self-hosted Kimi K3 weights can be used, modified and commercialised, and they run entirely on the deployer's own infrastructure, so data stays with the deployer. The custom, non-OSI licence adds a $20M Model-as-a-Service separate-agreement gate on top of the branding threshold. The training corpus, code and compute are closed. Most consequentially, the model ships with no safety tuning or guard while its own report shows it performing offensive-cyber tasks that frontier labs refuse. Self-host behind a full external control stack, avoid the train-by-default hosted API where data control matters, and confirm the LICENSE and the Model-as-a-Service threshold before commercial deployment.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The Kimi K3 model card on the verified moonshotai Hugging Face org lists a 2.8T total / 104B active MoE (16 of 896 experts), 1M context, MXFP4 weights / MXFP8 activations, native text+image+video, and serving on vLLM/SGLang/TokenSpeed.
Licenceread2026-11-19
The Kimi K3 License grants use, copying, modification, merging, publication, distribution, sublicensing and sale.
Technical_reportread2026-11-19
Moonshot's technical report "Kimi K3: Open Frontier Intelligence" (GitHub PDF) describes a 2.8T/104B MoE with Kimi Delta Attention and Per-Head Muon, and a context curriculum of 8K to 64K (pretrain) and 256K to 1M (cooldown).
Terms of serviceread2026-11-19
The platform.kimi.ai model-use agreement, dated "Last Updated: July 30th, 2026", allows user Content to be used to develop and improve the Services, with opt-out only via an enterprise arrangement on request.
Privacy Policyread2026-11-19
The platform.kimi.ai privacy policy states that the hosted service trains on user prompts, audio, images, videos and files by default ("helps us optimize our models").
Documentationread2026-11-19
The hosted Kimi API Platform model list calls kimi-k3.1-pro "the first model of the new K3.1 series and Kimi's most capable flagship to date" and kimi-k3 "Kimi's previous-generation flagship"; both are live.
Third-party analysisread2026-07-28
Tom's Hardware (reputable coverage) reports on the 2.8T Kimi K3 release.
Third-party analysisunverified2026-07-28
No public EU AI Act training-content summary, copyright policy or provider documentation package exists for Kimi K3, and the training corpus, training code and training-compute figures are not released.