Model
Kimi K3
Publisher
Moonshot AI (China)
Licence
Kimi K3 License (custom, modified-MIT with commercial thresholds)
You substantially use, modify and commercialise the self-hosted Kimi K3 weights and run them entirely on your own infrastructure, so your data stays yours - but ownership is only partial. 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/compute are closed, and, 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, keep off the train-by-default hosted API if data control matters, and confirm the LICENSE and the Model-as-a-Service threshold before commercial deployment.
Do you really own it?
Partial
none·limited·partial·substantial·full
Analytical input: AOI D · 53.2/100
The four ownership factors
Floor-weighted, not averaged. Reliability is weak, and the weakest factor sets the ceiling, so the verdict stays partial however strong the rest.
1
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
ModerateThe custom Kimi K3 License grants use, modification, redistribution and sale of the downloadable weights, but it is a community-style licence with real 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'. Commercially usable for most adopters, gated for large Model-as-a-Service operators - open-weight with conditions, 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/Partial, but the training data and training code are undisclosed (the report does not even state the training-token count) and evaluation is only partially reproducible.
Legal2/5how permissive and clean the licence is for real commercial usePermissive-enough for many uses but with material caveats: the custom "Kimi K3 License" is not OSI-certified and adds a $20M/12-month Model-as-a-Service separate-agreement gate on top of the 100M-MAU / $20M-revenue branding threshold, so the commercial grant carries real scale-linked conditions.
2
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateYou can inspect the weights and read a detailed technical report, but the training corpus, the training code and even the training-token and FLOP figures are closed, so you cannot see how it was made or reproduce it.
How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelDistributed from the verified moonshotai org on Hugging Face as native MXFP4/MXFP8 weights (not pickle) with the canonical source clear and no malicious-mirror incident on record (checklist ~4/8).
Governance3/5how accountable and well-documented the publisher isActive, accountable publisher (Moonshot AI) with a detailed technical report and a verified hub presence, but the report is a GitHub PDF rather than an archival venue, there is no documented vulnerability-disclosure or deprecation policy, and there is no EU Code of Practice signature.
3
ReliabilityIs it reliable and good enough for the job?
WeakCapability is frontier-competitive, but safety scores 1: no safety tuning, guard, or refusal training, and the report documents offensive-cyber capability the model does not refuse - a model presented as deployable whose safety is left entirely to the adopter. Under the ownership rule, safety <= 1 for a deployable model makes reliability weak regardless of benchmark strength.
How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionRuns on mainstream serving stacks (vLLM, SGLang, TokenSpeed) with native MXFP4 quantization and community quants available, so there is a practical path to run it - but the 2.8T-parameter scale is a heavy operational burden: minimum viable deployment is a multi-GPU / multi-node cluster (~1.4 TB of weights), there are no small variants, exact hardware guidance is sparse, and the documented stack is narrower than the K2 lineage (no KTransformers / TensorRT-LLM listed).
Safety1/5whether misuse risks are evaluated and guardrails are providedNo safety-tuned release, companion guard model, refusal training, or content filtering is described (model card and report), and the technical report's section 6.2.2 documents offensive-cyber capability - Tier-1 vulnerability discovery and Tier-2 end-to-end exploit development - explicitly noting that frontier Anthropic and OpenAI models refuse these tasks while K3 does not.
4
Doesn't extract your dataDoes running it keep your knowledge and data yours?
StrongSelf-hosted, the weights run entirely on your own infrastructure with no telemetry or licence claw-back, so your data stays yours. The hosted platform.kimi.ai service is the opposite: its terms and privacy policy train on user prompts, images, video and files BY DEFAULT (opt-out only via an enterprise arrangement on request, Singapore law / SIAC), so keep to the weights if data control is the point. The practical constraint on self-hosting is the ~1.4 TB footprint that forces multi-node serving.
How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
How the AOI score is computed
The seven dimensions above, each scored 0 to 5, weighted and summed to the 0 to 100 headline. The score is the analytical input behind the ownership verdict, not the verdict itself.
DimensionScoreWeightPoints
Openness3/50.1810.8
Provenance3/50.169.6
Legal2/50.166.4
Safety1/50.163.2
Performance4/50.1411.2
Operational3/50.127.2
Governance3/50.084.8
HeadlineD · 53.2/100
Sources
Every rating traces to a primary document. Read means the text was verified; unverified means it is known to exist but has not yet been read.
DocumentWhat it grounds
Model cardread2026-07-28
Kimi K3 model card on the verified moonshotai Hugging Face org: 2.8T total / 104B active MoE (16 of 896 experts), 1M context, MXFP4 weights / MXFP8 activations, native text+image+video, serving on vLLM/SGLang/TokenSpeed; no training data/code/tokens or safety tuning disclosed.
Licenceread2026-07-28
Kimi K3 License, read verbatim via the raw mirror: grant to use/copy/modify/merge/ publish/distribute/sublicense/sell, plus a Model-as-a-Service separate-agreement requirement over $20M aggregate revenue in any consecutive 12 months, plus a 100M-MAU or $20M-monthly- revenue "display Kimi K3" branding clause; internal-use and official/certified-partner exemptions; not OSI-certified.
Technical_reportread2026-07-28
Moonshot's technical report "Kimi K3: Open Frontier Intelligence" (GitHub PDF), read: 2.8T/104B MoE with Kimi Delta Attention and Per-Head Muon; context curriculum 8K to 64K (pretrain) and 256K to 1M (cooldown); no total training-token count and no absolute pretraining-FLOP figure stated; section 6.2.2 documents offensive-cyber capability (Tier-1 vulnerability discovery, Tier-2 exploit development) and notes Anthropic/OpenAI models refuse such tasks while K3 does not.
Terms of serviceread2026-07-28
platform.kimi.ai model-use agreement, read: user Content may be used to develop and improve the Services, with opt-out only via an enterprise arrangement on request; governed by the laws of Singapore with SIAC arbitration.
Privacy Policyread2026-07-28
platform.kimi.ai privacy policy, read: the hosted service trains on user prompts, audio, images, videos and files by default ("helps us optimize our models"); controller MOONSHOT AI PTE.
Third-party analysisread2026-07-28
Reputable coverage (Tom's Hardware) of the 2.8T Kimi K3 release: architecture (16/896 experts), MXFP4/MXFP8 QAT, K2 comparison, 1st place on the Frontend Code Arena, and positioning behind only Claude Fable 5 and GPT-5.6 Sol overall while ahead of other open models on coding and agentic benchmarks.
Third-party analysisunverified2026-07-28
No public EU AI Act training-content summary, copyright policy, or provider documentation package is published for Kimi K3, and the training corpus, training code and training-compute figures are not released.