Audience: Engineers deploying Kimi K3, compliance officers evaluating Chinese-origin open-weight models, security teams assessing supply chain risk for large-scale MoE deployments.

EU AI Act enforcement is live. GPAI obligations enforce August 2, 2026. K3 is a general-purpose AI model under Art. 51. Deployers of GPAI models bear compliance obligations regardless of the model's country of origin. The governance pattern is model-agnostic. SWT3 does not know or care who built the model. It records what happened and produces verifiable evidence.

Contents

1. K3 Model Specifications 2. The Governance Gap 3. Chinese-Origin Model Compliance 4. SWT3 Procedures for K3 5. Quick Start 6. Framework Mapping 7. Open-Weight Model Comparison 8. Quick Reference for Examiners 9. References

1. K3 Model Specifications

Attribute Value
Parameter count2.8 trillion (MoE)
Active parameters~400B per inference (estimated)
Expert count896
Context window1 million tokens
ArchitectureMixture of Experts (MoE)
LicenseApache 2.0
ProviderMoonshot AI (Beijing, China)
Open weights releaseJuly 27, 2026
API compatibilityOpenAI-compatible
ServingvLLM, SGLang, native API

K3 is the world's first open 2.8T model. At 896 experts, it represents the largest openly available mixture-of-experts architecture. The Apache 2.0 license allows unrestricted commercial use. Open weights ship with zero built-in governance controls.

2. The Governance Gap

SWT3 closes this gap. Every inference is recorded in a cryptographically signed SWT3 Witness Anchor. The anchor is independently verifiable and exists outside of Moonshot AI's infrastructure. You own the evidence chain.

3. Chinese-Origin Model Compliance

Country of origin does not change governance obligations. A model built in Beijing, deployed in Frankfurt, serving customers in New York has obligations under EU AI Act, US executive orders, and potentially Chinese regulations simultaneously.

GPAI Classification

K3 qualifies as a general-purpose AI model under EU AI Act Art. 51. It may also qualify as a GPAI model with systemic risk under Art. 51(2) given its 2.8T parameter count and general-purpose capabilities. Systemic risk classification triggers additional obligations under Art. 55, including adversarial testing and incident reporting.

Export Control Considerations

Deployers should verify that their use case does not trigger export control restrictions (EAR, Entity List). SWT3 provides evidence of deployment context but does not provide legal analysis. Consult export control counsel before deploying K3 in regulated industries or to restricted end users.

Chinese Regulatory Context

China's Interim Measures for Generative AI (effective July 2025) require service providers operating within China to register models. Deployers outside China using open-weight K3 weights are not directly subject to these registration requirements, but awareness of the regulatory landscape informs supply chain risk assessment.

The Model-Agnostic Principle

The governance pattern is model-agnostic. SWT3 does not know or care who built the model. It records what happened and produces verifiable evidence. The same SWT3 Witness Anchor format applies to K3 deployed in Frankfurt as it does to Claude deployed in us-east-1.

4. SWT3 Procedures for K3

AI-INF.1

Inference Provenance

K3 context: At 896 experts with dynamic routing, each inference activates a different subset of the model. Without provenance records, there is no way to reconstruct which expert combination processed a given request. AI-INF.1 records the model identifier, input/output hashes, latency, and serving endpoint for every inference.

Assessor Tip

AI-INF.1 anchors prove that a specific inference occurred at a specific time through a specific K3 deployment. The anchor exists independently of Moonshot AI's infrastructure. Check that model_id in each anchor resolves to a known K3 version identifier.

AI-GRD.1

Guardrail Enforcement

K3 context: K3 ships without built-in safety filters. Unlike Claude or GPT-4, there is no provider-side content filtering. Deployers must implement their own guardrails (Llama Guard, NeMo Guardrails, custom classifiers) and prove they were active at inference time.

Assessor Tip

AI-GRD.1 evidence shows guardrail status per inference. A pattern of guardrailsActive = false indicates unprotected inferences from a 2.8T model with no built-in safety layer. This is a finding under EU AI Act Art. 9 for any high-risk use case.

AI-MDL.1

Model Integrity Hash

K3 context: 2.8T parameters distributed across potentially hundreds of shard files. The weights hash proves which exact version is deployed. Any modification (fine-tuning, quantization, or tampering) produces a different hash. This is especially critical for models distributed through third-party mirrors where provenance is harder to verify.

Assessor Tip

Compare the recorded weights hash against the published hash from Moonshot AI or Hugging Face. For a 2.8T model, the integrity chain is especially important given the infrastructure required to host it. A mismatch indicates either a modified model or an undocumented quantization.

AI-SUPPLY.1

Supply Chain Integrity

K3 context: The supply chain for K3 includes model weights (Moonshot AI, Hugging Face mirrors), serving framework (vLLM, SGLang), compute dependencies (CUDA, PyTorch), and container images. For Chinese-origin models, supply chain documentation is particularly important for demonstrating due diligence to auditors and security reviewers.

Assessor Tip

AI-SUPPLY.1 anchors document the full provenance chain: weights source, framework version, dependency manifest. Compare against published checksums from Moonshot AI. Any gap in the chain is a finding under NIST AI RMF GOVERN 1.4 and EU AI Act Art. 53(1)(a).

AI-DUALUSE.1

Dual-Use Risk Assessment

K3 context: A 2.8T model with general-purpose capabilities and no built-in restrictions has inherent dual-use potential. AI-DUALUSE.1 records the deployer's risk assessment for dual-use scenarios, documenting what use cases are permitted, what restrictions are in place, and what monitoring exists for misuse detection.

Assessor Tip

AI-DUALUSE.1 anchors prove that the deployer assessed dual-use risk before deployment. Required for GPAI models with systemic risk under EU AI Act Art. 55(1)(a). Absence of this record is an automatic finding for any K3 deployment assessed under EU AI Act obligations.

5. Quick Start

OpenRouter (managed endpoint)

from openai import OpenAI
from swt3_ai import SWT3Witness

# K3 via OpenRouter (managed endpoint)
client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="your_openrouter_key"
)
witness = SWT3Witness(
    api_key="axm_live_your_key",
    tenant_id="your_tenant_id",
    agent_id="k3-openrouter-prod"
)
wrapped = witness.wrap(client)

response = wrapped.chat.completions.create(
    model="moonshot/kimi-k3",
    messages=[{"role": "user", "content": "Analyze this regulatory filing."}]
)

vLLM self-hosted (sovereign deployment)

from openai import OpenAI
from swt3_ai import SWT3Witness

# K3 self-hosted via vLLM
client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")
witness = SWT3Witness(
    api_key="axm_live_your_key",
    tenant_id="your_tenant_id",
    agent_id="k3-vllm-sovereign"
)
wrapped = witness.wrap(client)

response = wrapped.chat.completions.create(
    model="moonshot-ai/Kimi-K3",
    messages=[{"role": "user", "content": "Review this compliance report."}]
)

Both patterns produce identical SWT3 Witness Anchors. The evidence chain is the same whether K3 runs on OpenRouter infrastructure or your own GPU cluster. Install the SDK: pip install swt3-ai. Get an API key at /signup.

6. Framework Mapping

Framework Relevant Requirements SWT3 Evidence
EU AI Act (GPAI) Art. 51 GPAI classification, Art. 53 technical documentation, Art. 55 systemic risk AI-INF.1, AI-MDL.1, AI-DUALUSE.1
EO 14110 (superseded June 2026) Dual-use foundation model reporting, safety evaluations AI-DUALUSE.1, AI-GRD.1, AI-INF.1
NIST AI RMF MAP 1.5 risk identification, MEASURE 2.6 bias, GOVERN 1.1 accountability AI-GRD.1, AI-INF.1, AI-SUPPLY.1
ISO 42001 6.1 risk assessment, 8.4 data management, 9.1 monitoring AI-INF.1, AI-MDL.1, AI-GRD.1

Explore all framework crosswalks in the crosswalk explorer, including EU AI Act GPAI obligations and NIST AI RMF procedure mappings.

7. Open-Weight Model Comparison

Model Provider Parameters License Governance Controls SWT3 Integration
K3 Moonshot AI 2.8T MoE Apache 2.0 None wrap(OpenAI(...))
Llama 4 Meta 109B-400B Llama License None wrap(OpenAI(...))
DeepSeek V3 DeepSeek 685B MoE MIT None wrap(OpenAI(...))
Inkling Thinking Machines 975B MoE Apache 2.0 None wrap(OpenAI(...))

All open-weight models produce identical SWT3 Witness Anchors. The governance pattern is model-agnostic.

8. Quick Reference for Examiners

Examiner Question SWT3 Evidence Procedure
What model version processed this request? Anchor model_id field, cross-referenced to weights hash AI-INF.1, AI-MDL.1
Were safety controls active during inference? Per-inference guardrailsActive flag in anchor payload AI-GRD.1
Has the model been modified from published weights? SHA-256 weights hash compared to Moonshot AI published checksum AI-MDL.1
Where did the model weights come from? Supply chain manifest in AI-SUPPLY.1 anchor: source URL, mirror hash, framework version AI-SUPPLY.1
Was dual-use risk assessed before deployment? AI-DUALUSE.1 anchor with timestamp predating first inference anchor AI-DUALUSE.1
Can you reconstruct the inference timeline? Ordered anchor sequence by epoch timestamp, each with SHA-256 fingerprint AI-INF.1
What is the GPAI evidence package? Export from /api/v1/ai-witness/export: full anchor chain, procedure map, framework crosswalk All

9. References