2.8조 파라미터 · 오픈웨이트 · MoonEP/FlashKDA/AgentEnv · 라이선스 · 미중 AI 분쟁
결론부터 말하면 엄격한 정의로는 오픈소스가 아닙니다 — Moonshot AI도 그렇게 주장하지 않습니다. 2026년 7월 27일 밤, Kimi K3 전체 가중치와 기술 보고서를 공개하고 MoonEP·FlashKDA·AgentEnv도 오픈소스화했습니다. 오픈웨이트 vs 오픈소스, 라이선스 상업 게이트, GPT-5.6 Sol·Claude Fable 5 비교가 필요하다면 타임라인, 아키텍처, 벤치마크 표, 6단계 Runbook, 라이선스 분석으로 실행 가능한 결론을 제공합니다.데이터 기준: 2026-07-28
By the Open Source Initiative definition, a genuinely open-source model requires public training data, public training code, and a reproducible pipeline — not just trained weights. Kimi K3 ships weights, a technical report, and training-adjacent infrastructure tools, but not training data or full training code. Moonshot consistently calls this an open-weight release, never open source.
The license has also tightened versus the K2 family. It is no longer branded Modified MIT — it is a fully custom document with two commercial thresholds:
Model-as-a-Service revenue gate: If you or affiliates run a MaaS business on K3 generating more than $20 million in aggregate trailing-12-month revenue, you must negotiate a separate commercial agreement with Moonshot AI.
Attribution gate: If your product exceeds 100 million MAU or $20 million monthly revenue, you must prominently display Kimi K3 in the UI.
July 16 API launch: K3 went live on kimi.com and the API; weights were not yet public.
July 22–23 distillation dispute: White House adviser Michael Kratsios accused Moonshot of industrial distillation against Anthropic Fable; Treasury Secretary Scott Bessent floated sanctions.
July 27 full release: ~1.56TB weights hit Hugging Face trending #1 within 30 minutes; MoonEP and AgentEnv open-sourced (FlashKDA was already out).
Teams evaluating K3 often hit five blind spots before shipping:
Semantic confusion: Media says open source; Moonshot says open weight — training data and code are not public.
License traps: New MaaS revenue gate did not exist in the K2 Modified MIT era.
Self-hosting fantasy: 2.8T parameters need 64+ accelerator supernodes — not a prosumer GPU rig.
Benchmark trust: Vendor numbers diverge; cross-check Vals AI and Artificial Analysis independently.
Cost vs capability: K3 is the open-weight capability ceiling at ~$0.95/task, but GLM-5.2 runs ~$0.47/task.
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion |
| Active parameters | ~104 billion |
| Architecture | Mixture-of-Experts (MoE) |
| Experts | 896 routed, 16 activated per token, plus shared experts |
| Attention | Kimi Delta Attention (KDA) + Gated MLA |
| Context window | 1,000,000 tokens |
| Multimodality | Native vision (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights, MXFP8 activations (QAT from SFT) |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom — open weight, not open source |
Standard Gated DeltaNet applies one scalar forgetting gate to an entire memory state. KDA uses channel-wise gating: every feature dimension gets its own decay rate. Implemented as chunkwise Diagonal-Plus-Low-Rank (DPLR), it keeps recurrence linear in time while staying hardware-efficient. K3 interleaves KDA with Gated MLA global-attention layers — the hybrid that supports 1M context while keeping KV cache small.
Uniform residual accumulation dilutes early-layer signal at depth. AttnRes replaces it with selective, input-dependent aggregation across preceding layers — one RMSNorm and one pseudo-query vector per layer, roughly 25% training efficiency gain for under 2% added cost.
Per-Head Muon optimizes attention heads independently. The MoE layer routes 896 experts, activates 16 per token (~1.8% sparsity), and uses Quantile Balancing plus MoonEP to prove a mathematical upper bound on redundant experts per rank.
Kimi K3 rebuilt three long-standing defaults — attention, residual connections, and the optimizer — rather than simply scaling an existing recipe.
Confirm compliance boundaries: Read the custom license. MaaS revenue over $20M trailing 12 months requires a separate Moonshot agreement; 100M+ MAU or $20M+ monthly revenue requires prominent Kimi K3 attribution.
Get an API key: Create a project at platform.moonshot.ai and verify balance and rate limits.
Configure the OpenAI-compatible SDK: Set base_url=https://api.moonshot.ai/v1, model ID kimi-k3 (see code below).
Optional OpenRouter path: Model ID moonshotai/kimi-k3 — seven providers already host it, mostly at official pricing.
Optimize cache hit rate: Mooncake disaggregated serving reportedly exceeds 90% cache hits on coding workloads — real input cost lands near $0.30/M, not the $3 headline.
Self-hosting feasibility check: Moonshot recommends 64+ accelerator supernodes. Individual developers should use API or OpenRouter, not local weights.
from openai import OpenAI
client = OpenAI(
api_key="your_moonshot_api_key",
base_url="https://api.moonshot.ai/v1"
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Analyze this code..."}]
)
| Technology | Role | Key numbers |
|---|---|---|
| MoonEP | MoE communication at extreme scale | Duplicates overloaded experts for equal token counts; proves redundant-expert upper bound per rank |
| FlashKDA | CUTLASS KDA kernels | 1.72–2.22× faster prefill on H20 vs flash-linear-attention; drop-in via chunk_kda |
| AgentEnv | Firecracker microVM agent sandbox | Checkpoint ~133ms, resume ~49ms, up to 6.5× memory overcommit (Moonshot-reported) |
| Model | Score | Release |
|---|---|---|
| Claude Opus 5 | 97% | 2026-07-24 |
| GPT-5.6 Sol | 96.2% | 2026-07-09 |
| Claude Fable 5 | 95% | 2026-06-09 |
| Kimi K3 | 93.4% | 2026-07-16 |
| Qwen3.7-Max | 79.4% | 2026-05-19 |
| DeepSeek-V4 | 76.2% | 2026-04-23 |
Artificial Analysis Intelligence Index (max reasoning): Kimi K3 scores ~57 — #3 overall, #1 among open-weight models, behind Claude Fable 5 (60) and GPT-5.6 Sol (59). ~$0.95/task vs Fable 5 ~$2.40 (~60% cheaper) but above GLM-5.2 ~$0.47. K3 ranks #1 on Arena.ai Frontend Code Arena.
| Token type | Price per million |
|---|---|
| Input (cache hit) | $0.30 |
| Input (cache miss) | $3.00 |
| Output (incl. reasoning) | $15.00 |
License gates matter if you resell inference at scale. For nearly every startup or mid-size company, neither threshold is practical — you can ship commercial products on K3 today. The clause that bites is the MaaS revenue gate, and only if you compete directly with Moonshot API.
The release landed during WAIC 2026 and days after the US-China distillation dispute escalated. China's Ministry of Commerce pushed back July 28, accusing Washington of AI hegemonism. Three days later Alibaba — a Moonshot investor — unveiled Qwen3.8-Max-Preview (2.4T parameters), pushing the race into what some call the 3T club.
In short: Kimi K3 is the capability ceiling of the open-weight tier, not the value pick.
API로 K3 추론은 가능하지만 iOS 서명 체인, Xcode 빌드, Metal 워크로드, 24/7 CI에는 실제 macOS 노드가 필요합니다. 프로덕션 iOS CI/CD와 AI Agent 자동화에는 VpsMesh Mac Mini 클라우드 대여가 일반적으로 더 나은 선택입니다. Mac Mini M4 대여 가격과 고객 센터를 참고하세요.
Sources: Moonshot AI official blog, Hugging Face, GitHub moonshotai/FlashKDA, Vals AI SWE-bench, Artificial Analysis, VentureBeat, Simon Willison. Verify against official data before production decisions.
No, not by the OSI strict definition. It is open-weight: trained weights, technical report, and some infrastructure tools are public, but training data and full training code are not. Moonshot uses open weight, never open source, in its own materials.
Yes, for the vast majority of use cases. Restrictions apply only if you run a Model-as-a-Service business on K3 generating over $20 million in trailing 12-month revenue (requires a separate Moonshot agreement), or if you exceed 100 million MAU or $20 million monthly revenue (requires displaying Kimi K3 in your UI).
Moonshot은 64개 이상 액셀러레이터 슈퍼노드 구성을 권장합니다. 개인 개발자는 공식 API 또는 OpenRouter를 사용하세요. 24/7 빌드 환경은 Mac Mini M4 대여 가격을 확인하세요.
It trails Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 on independent SWE-bench Verified (93.4% vs 95–97%) and ranks #3 on the Artificial Analysis Intelligence Index, but it is the strongest open-weight model available, ahead of GLM-5.2 and DeepSeek V4 Pro.
K3 has roughly 3× the parameters of K2.5, adds Attention Residuals and Per-Head Muon, and significantly expands context and multimodal capability. Its license adds a new Model-as-a-Service revenue threshold not present in K2. More at our help center.