KIMI K3
OPEN_
SOURCE_
ИЛИ_НЕТ.

Kimi K3 open-weight MoE модель 2,8 триллиона параметров

Короткий ответ: нет, не по строгому определению — и Moonshot AI сам это не утверждает. 27 июля 2026 Moonshot опубликовала полные веса, technical report и три infra-технологии: MoonEP, FlashKDA и AgentEnv. Pain point: «open source» и «open weight» — не одно и то же; custom license содержит два commercial gate, которые большинство обзоров игнорирует. Вывод: K3 — capability ceiling open-weight tier (AA Index #3 globally), не value pick. Структура: license clarity → specs → KDA/AttnRes architecture → infra stack → benchmarks → pricing/self-hosting → geopolitical context → 5-step integration → FAQ.

1. Pain points: почему 27 июля меняет оценку

1) Terminology trap: Moonshot последовательно говорит «open weight», никогда «open source» — training data и полный training code не публичны. 2) Два license gate: Model-as-a-Service revenue выше $20M за 12 месяцев или продукты с 100M+ MAU / $20M monthly revenue активируют дополнительные условия. 3) Self-hosting нереалистичен: 1,56 TB download, 64+ accelerator supernode recommended — API или OpenRouter = практичный путь. 4) Не самый дешёвый: ~$0,95/task на Intelligence Index vs GLM-5.2 ~$0,47. 5) Geopolitical overlay: US distillation allegations и counter-response Китая 28 июля добавляют compliance и reputational risk при выборе vendor.

2. Timeline: от API launch до full weights за 11 дней

ДатаСобытие
16 июляK3 запуск на kimi.com, Kimi Work, Kimi Code и API — веса withheld
17 июляIndustry architecture analysis; state media называет largest open model по parameter count
22–23 июляUS officials обвиняют Moonshot в industrial-scale distillation из Anthropic Fable
27 июля, 23:00Full weights, technical report, MoonEP и AgentEnv open-sourced (FlashKDA уже public)
28 июляMinistry of Commerce Китая отвечает; domestic media покрывает release details

1,56 TB Hugging Face release занял #1 trending за 30 минут — всего 11 дней после API-only launch.

3. Kimi K3 at a glance

SpecValue
Total parameters2,8 trillion
Active parameters~104 billion
ArchitectureMixture-of-Experts (MoE)
Experts896 routed, 16 activated per token + shared experts
AttentionKimi Delta Attention (KDA) + Gated MLA
Context window1 000 000 tokens
MultimodalityNative vision (ViT-V2, 27 layers)
Weight formatMXFP4 weights, MXFP8 activations (QAT from SFT stage)
Download size~1,56 TB (Hugging Face)
LicenseCustom — «open weight», not «open source»

4. Architecture: KDA, Attention Residuals, Per-Head Muon

Kimi Delta Attention (KDA)

Standard Gated DeltaNet применяет single scalar forgetting gate. KDA заменяет его channel-wise gating — каждая feature dimension получает свой decay rate. Реализовано как chunkwise DPLR recurrence, interleaved с Gated MLA global-attention layers. Hybrid design поддерживает 1M-token context при низком KV cache footprint.

Attention Residuals (AttnRes)

Uniform residual accumulation размывает early-layer information на depth. AttnRes использует selective, input-dependent aggregation across preceding layers — один RMSNorm и один pseudo-query vector per layer, ~25% higher training efficiency при <2% added parameter cost.

Stable LatentMoE

896 experts, 16 activated per token (~1,8% sparsity). Quantile Balancing + MoonEP address expert load imbalance с proven upper bound на redundant experts per rank.

5. Три infrastructure release

TechnologyRoleKey numbers
MoonEPMoE communication at extreme scaleDuplicates overloaded experts temporarily; proves redundant-expert upper bound per rank
FlashKDACUTLASS-based KDA kernel1,72×–2,22× faster prefill on H20 vs flash-linear-attention baseline; drop-in via chunk_kda
AgentEnvFirecracker microVM agent sandbox (with KVCache.ai)Checkpoint 133 ms, resume 49 ms, up to 6,5× memory overcommit (vendor-reported, not independently verified)

6. Benchmarks vs GPT-5.6, Claude и GLM-5.2

SWE-bench Verified (Vals AI independent evaluation, July 2026)

ModelScoreRelease
Claude Opus 597%2026-07-24
GPT-5.6 Sol96,2%2026-07-09
Claude Fable 595%2026-06-09
Kimi K393,4%2026-07-16
Qwen3.7-Max79,4%2026-05-19
DeepSeek-V476,2%2026-04-23

Artificial Analysis Intelligence Index (max reasoning): Claude Fable 5 (60) > GPT-5.6 Sol (59) > Kimi K3 (~57, #3 overall, #1 open-weight) > GLM-5.2 (51) > DeepSeek V4 Pro (44).

~$0,95/task на Intelligence Index — ~60% cheaper vs Claude Fable 5 (~$2,40), но дороже GLM-5.2 (~$0,47). Capability ceiling open-weight tier, не value pick. K3 #1 на Arena.ai Frontend Code Arena.

7. License: два commercial gate

  1. MaaS revenue gate: Model-as-a-Service businesses с aggregate revenue >$20 million за trailing 12 months требуют separate commercial agreement с Moonshot.
  2. Attribution gate: Products с >100 million MAU или >$20 million monthly revenue должны prominently display «Kimi K3» в UI.

Для nearly every startup и mid-size company ни один gate не practical concern. Первый clause relevant только если business model = reselling K3 inference at scale в direct competition с Moonshot API.

8. API pricing и self-hosting

Token typePrice per million tokens
Input (cache hit)$0,30
Input (cache miss)$3,00
Output (incl. reasoning trace)$15,00

Endpoint: https://api.moonshot.ai/v1, model ID kimi-k3, OpenAI SDK compatible. Mooncake disaggregated serving держит 90%+ cache hit rates на coding workloads — real-world input cost ближе к $0,30 чем $3,00. Self-hosting требует 64+ accelerator supernode; OpenRouter уже lists seven providers.

9. Пять шагов интеграции K3

  1. Pick your path: API vs third-party hosting vs self-deploy — 99% teams выбирают первые два.
  2. Audit the license: Check MaaS revenue и MAU thresholds против business model.
  3. Wire OpenAI-compatible clients: Cursor, OpenClaw или custom agents — change base URL и API key only.
  4. Exploit cache hits: Repeated code context на Mooncake architecture drives effective cost к $0,30 tier.
  5. Build fallback chain: Route hard tasks к K3, daily traffic к GLM-5.2 или DeepSeek V4 Flash для cost control.

10. Deep dive: WAIC 2026 и гонка «3T club»

K3 release попал на WAIC 2026 и days after US-China distillation dispute escalated — White House adviser Michael Kratsios обвинил Moonshot в industrial-scale distillation из Anthropic Fable; Ministry of Commerce Китая push back 28 июля, accusing Washington of «AI hegemonism». Release full weights, technical report и three infrastructure technologies — deliberate signal: Moonshot betting на engineering transparency как clearest answer на questions about methods.

Three days later Alibaba — Moonshot investor — unveiled Qwen3.8-Max-Preview at 2,4 trillion parameters, widely read as direct response. China open-weight race entered «3T club» — parameter scale, architectural innovation и infra open-sourcing moving in lockstep.

Для Mac devs self-hosting K3 not realistic, API access trivial. Local MLX fallbacks с quantized GLM-5.2-class models viable на Apple Silicon. K3 real value: proving open-weight models can approach closed frontier capability, shifting pricing и benchmark baselines для entire ecosystem.

11. FAQ

Is Kimi K3 open source? No, not by OSI definition. Open-weight: trained weights, technical report и some infra tools public; training data и full training code — not.

Can I use it commercially? Yes, для vast majority use cases. Restrictions apply only at MaaS revenue и scale thresholds выше.

How much hardware to self-host? Moonshot recommends 64+ accelerators на supernode. Most developers — official API или OpenRouter.

How does K3 differ from K2? ~3× K2.5 parameter count, adds AttnRes и Per-Head Muon, expands context и multimodal capability; license adds new MaaS revenue threshold.

12. Closing: API access + Mac local fallback architecture

Windows или Linux cloud box может вызвать Kimi K3 API, но проигрывает в Cursor toolchain integration, OpenClaw 24/7 agent hosting, local MLX quantized fallbacks и graphics/multimedia workflows vs Apple Silicon Macs. K3 self-hosting needs 64-GPU supernode; API access — только base URL change. Если нужны local GLM-5.2 offline backup, long-context agent offload и stable always-on operation, three-tier stack optimal: local MLX для daily quantized open models; Kimi K3 API для hard coding и reasoning; MACGPU remote Mac nodes для OpenClaw/Hermes residency и unified-memory-heavy long-context workloads. Zero PCIe bottleneck на KV-heavy agent sessions vs discrete GPU + system RAM split — Metal compute stack на M4 Max/M4 Ultra держит sustained tokens/sec на 70B-class MLX quants без HBM pressure. Remote node съедает agent memory overcommit, пока main machine остаётся stable для everyday dev; track Metal throughput numbers при 1M-token batch jobs перед prod deploy.