Alpha
Contact
Implement · Kimi K3

Can you run it?

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

This page is a projection of the one entry record, the Reliability and Data control factors that Implement 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

Install & run

Download the native MXFP4/MXFP8 weights from the verified moonshotai org on Hugging Face (huggingface.co/moonshotai/Kimi-K3) and serve them with vLLM, SGLang or TokenSpeed. Pin the exact revision hash and verify per-file checksums before loading.

Hardware & VRAM requirements

Kimi K3 is a 2.8T-parameter MoE (104B active). Even at native MXFP4 the full weights are about 1.4 TB, so serving needs a multi-GPU or multi-node cluster; there is no single-GPU or laptop path. The model card does not state exact per-GPU VRAM, so size against the cluster's aggregate memory and interconnect.

Serving stacks

The card documents vLLM, SGLang and TokenSpeed. KTransformers and TensorRT-LLM are not listed, so the supported stack is narrower than the Kimi K2 lineage. MXFP4/MXFP8 kernels are version-sensitive, so confirm current engine support and versions before committing.

Safe-deployment controls & Deployment Ceiling

Deployment Ceiling: T2 (customer-facing, human-reviewed), conditional. The weights ship with no safety tuning, guard model or refusal training, so the entire control stack must be supplied externally: a companion guard or classifier model, input and output guardrails, prompt-injection defences and conservative decoding.

The technical report documents offensive-cyber capability the model does not refuse. Sandbox and tightly bound any agentic, tool-use or code-execution deployment. T3+ (autonomous) is cautioned against without controls beyond what an open model can currently justify.

Complete the pre-deployment gate in the safe-deployment playbook, pin and checksum the checkpoint, and red-team for the specific use case before any exposed deployment.

Available quantizations

The weights are natively quantization-aware-trained (MXFP4 weights, MXFP8 activations), so the canonical checkpoint is already a low-precision artifact. Community MXFP4/GGUF quants make the 2.8T model more tractable but are community artifacts; verify each one.

Fine-tuning & adaptation

The downloadable weights and a broadly commercial licence support SFT and LoRA adaptation. Training data and code are closed, so there is no from-scratch reproduction path. Full fine-tuning at 2.8T scale is very resource-intensive, so parameter-efficient adaptation is the practical route without frontier-scale infrastructure.

API / OpenAI-compatible integration

Serve behind the OpenAI-compatible endpoint of vLLM or SGLang for drop-in client integration. Kimi K3 is natively multimodal, so text, image and video inputs and tool calls map onto the serving layer's APIs. Confirm the exact multimodal request format against the serving engine's documentation.

How this scores

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

3

ReliabilityIs it reliable and good enough for the job?

Weak

Capability is frontier-competitive, but safety scores 1: no safety tuning, guard or refusal training exists, and the report documents offensive-cyber capability the model does not refuse. The model is presented as deployable while 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 productionK3 runs on mainstream serving stacks (vLLM, SGLang, TokenSpeed) with native MXFP4 quantization and community quants, so a practical path exists.
Safety1/5whether misuse risks are evaluated and guardrails are providedNo safety-tuned release, companion guard model, refusal training or content filtering is described in the model card or report.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted, the weights run entirely on the deployer's own infrastructure with no telemetry or licence claw-back, so data stays with the deployer. The hosted platform.kimi.ai service differs: its terms and privacy policy train on user prompts, images, video and files by default, with opt-out only via an enterprise arrangement on request (Singapore law / SIAC). Data control therefore requires the weights and not the hosted API. Self-hosting is constrained by the ~1.4 TB footprint, which 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.
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.