Alpha
Contact
Implement · Kimi K2

Can you run it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C · 62.8/100

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

Kimi K2 weights are block-FP8 safetensors on the verified moonshotai org. This is not a single-command local model: plan for a multi-GPU or multi-node serving stack, and pin an exact revision.

# vLLM (OpenAI/Anthropic-compatible server), tensor/pipeline parallel across a node/cluster
vllm serve moonshotai/Kimi-K2-Instruct --revision <sha> \
  --tensor-parallel-size 8 --trust-remote-code

# SGLang (alternative high-throughput engine)
python -m sglang.launch_server --model-path moonshotai/Kimi-K2-Instruct --revision <sha>

KTransformers (heavy CPU offload) suits constrained GPU memory, and TensorRT-LLM suits maximum throughput. Verify per-file checksums after download.

Hardware & VRAM requirements

Kimi K2 is a 1T-parameter MoE (32B active). Even in block-FP8 the weights are about 1 TB, so serving needs a multi-GPU or multi-node cluster (e.g. a full 8×80-141 GB node or more) or heavy CPU offload via KTransformers. There are no small variants, so single-GPU or laptop deployment is not possible. Budget substantial additional headroom for the KV cache at the full 128K context.

Serving stacks

The publisher documents support for four high-end serving engines.

StackUse it for
vLLMHigh-throughput serving + OpenAI/Anthropic-compatible API
SGLangHigh-throughput serving, structured generation
KTransformersGPU-constrained deployment via CPU/GPU offload
TensorRT-LLMMaximum-throughput NVIDIA-optimised serving

There is no llama.cpp or Ollama single-box path at full precision. The model's scale gates deployment, which is why Operational Readiness scores 3.

Safe-deployment controls & Deployment Ceiling

Apply the safe-deployment playbook. The model-specific gate rows are:

  • Supply chain: pin the exact moonshotai revision hash and verify block-FP8 checksums. The weights are safetensors, so no pickle load occurs.
  • Model layer: deploy the Instruct variant, never the Base checkpoint, and set conservative decoding defaults.
  • Input/output layers: add guardrails and your own guard classifier for any customer-facing path, because Kimi K2 ships no companion guard model. Add prompt-injection defences and treat retrieved and tool content as untrusted, which matters given the native tool-use focus.
  • Behaviour: allow for China-aligned topic censorship and do not assume neutral coverage on politically sensitive prompts.
  • Compliance: self-assemble the copyright policy and training-content material, which are not available upstream.

Deployment Ceiling: T2 (customer-facing, human-reviewed), conditional. With the Instruct variant, input/output guardrails, a guard model and prompt-injection defences, Kimi K2 reaches customer-facing, human-reviewed use. T3+ (bounded autonomous), which its agentic positioning invites, also needs a deterministic action-policy engine, per-action human approval for irreversible operations and external guard tooling beyond what the release provides.

Available quantizations

Kimi K2 is distributed natively in block-FP8, which is the practical serving format. Community lower-bit and GGUF quants make the trillion-parameter model tractable on smaller clusters, but they are community artifacts, not a broad official quantization programme. Verify the source and test quality before relying on a third-party quant in production. Coverage is partial for that reason.

Fine-tuning & adaptation

A released Kimi-K2-Base checkpoint and the permissive modified MIT licence make adaptation possible, and standard SFT or LoRA is the practical path for most teams. Two limits keep coverage partial. The training data and code are not released, so continued pre-training and from-scratch reproduction are not possible. Full fine-tuning of a 1T-parameter MoE is resource-intensive, so most adopters are limited to parameter-efficient methods. Placing a derivative on the market carries EU AI Act provider implications (see Assess).

API / OpenAI-compatible integration

Serve Kimi K2 behind the OpenAI/Anthropic-compatible endpoint of vLLM or SGLang for drop-in integration with existing client code:

vllm serve moonshotai/Kimi-K2-Instruct --revision <sha> --tensor-parallel-size 8 --port 8000
# then point any OpenAI-compatible SDK at http://localhost:8000/v1

Kimi K2's native tool calling maps onto the function-calling API, so agentic clients and tool routers carry over unchanged and no model-specific client is required.

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?

Moderate

Runnable but heavy: dependable self-hosting of the trillion-parameter model is a multi-node exercise, and the safety tuning is light with no guard model.

How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionServing support is good (vLLM, SGLang, KTransformers, TensorRT-LLM) with an OpenAI/Anthropic-compatible API.
Safety3/5whether misuse risks are evaluated and guardrails are providedThe Instruct variant is safety-tuned and withstands casual jailbreaks.
4

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

Strong

Self-hosted, the weights run entirely on your own infrastructure with no telemetry or data clawback, so your data stays yours. The hosted platform.kimi.ai service differs. Its terms and privacy policy train on user prompts, audio, images, videos and files by default (opt-out only via an enterprise or separate written agreement, Singapore law and SIAC arbitration), so keep to the weights if data control matters. Self-hosting is constrained by the ~1 TB block-FP8 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 adoptionYou substantially use, modify and commercialise the self-hosted Kimi K2 weights under a near-MIT grant. The sole condition is the 100M-MAU / $20M-revenue trigger to display 'Kimi K2' in your UI. The weights run entirely on your own infrastructure, so your data stays yours. Ownership is partial because of transparency: the 15.5T-token corpus and training code are closed and the model carries China-aligned censorship that cannot be inspected. The weights and the hosted service differ: the platform.kimi.ai API trains on your content by default (opt-out only by enterprise agreement, Singapore law), so self-host if data control matters. Confirm the attribution trigger and the per-checkpoint LICENSE before shipping, and budget multi-node infrastructure to run it at all.

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-K2-Instruct model card on the verified moonshotai Hugging Face org describes block-FP8 safetensors, 1T total and 32B active parameters, a 128K context and the Muon optimiser.
Licenceread2026-11-19
The Kimi K2 LICENSE is standard MIT with one added clause, which opens "Our only modification part is that".
Documentationunverified2026-07-25
The moonshotai/Kimi-K2 repository documents the MoE architecture (384 experts, 8+1 selected), the MuonClip optimiser, block-FP8 weights and deployment on vLLM, SGLang, KTransformers and TensorRT-LLM with an OpenAI/Anthropic-compatible API.
Technical_reportread2026-07-25
Moonshot's arXiv technical report "Kimi K2: Open Agentic Intelligence" (2507.20534) describes a 1T-total / 32B-active MoE trained on 15.5T tokens with the MuonClip optimiser (Muon + QK-clip) and reports agentic and coding results (e.g.
Terms of serviceread2026-11-19
The platform.kimi.ai model-use agreement (page dated "Last Updated: July 30th, 2026") says user content is used to improve the services, with opt-out available only via an enterprise or separate written agreement.
Privacy Policyread2026-11-19
The platform.kimi.ai privacy policy says the hosted service trains on user prompts, audio, images, videos and files by default; the controller is MOONSHOT AI PTE.
Documentationread2026-11-19
The hosted Kimi API Platform notice says that from 2026-10-31 K2 series models are no longer available to newly registered users (existing users unaffected) and Moonshot plans to retire all K2 series models by the end of December 2026.
Third-party analysisunverified2026-07-25
On independent evaluation Kimi K2 is among the strongest open-weight models on agentic and coding benchmarks (e.g.
Third-party analysisunverified2026-07-25
Independent analysis notes Kimi K2, like other China-based open-weight models, applies China-aligned content filtering on politically sensitive topics.
Third-party analysisunverified2026-07-25
Kimi K2 Instruct is safety-tuned but with lighter alignment coverage than Western frontier labs and no companion guard model.
Third-party analysisunverified2026-07-25
No public EU AI Act training-content summary, copyright policy, or provider documentation package is published for Kimi K2, and the training corpus is not released.
Third-party analysisunverified2026-07-25
Kimi K2 is served across vLLM, SGLang, KTransformers and TensorRT-LLM, but its 1T-parameter scale requires multi-GPU / multi-node infrastructure even in block-FP8.