Kimi K3 — open weight или open source? Внутри 2,8T MoE от Moonshot AI
Вечером 27 июля 2026 Moonshot AI выкатила полные weights, technical report и три infra-стека для Kimi K3 — MoE на 2,8 трлн params, ~104B active, 1M-token context, native vision. Download ~1,56 TB на Hugging Face занял #1 trending за 30 минут.
TL;DR: не open source по OSI. Moonshot последовательно пишет open weight. Разбор для тех, кто будет деплоить: 11-day timeline API→weights; arch stack KDA/AttnRes/Per-Head Muon/LatentMoE; infra MoonEP/FlashKDA/AgentEnv; independent benchmarks (SWE-bench 93,4 %, AA index ~57); license gates ($20M MaaS, 100M MAU); API pricing и 64-GPU self-host floor. Контекст: deep-dive K3, distillation controversy.
01 Open weight vs open source: что реально в tarball
OSI definition требует public training data, training code, reproducible pipeline — не только checkpoint. K3 отдаёт weights + report + infra (MoonEP, FlashKDA, AgentEnv), но без training data и без full training codebase. В официальных материалах Moonshot — только open weight, zero mentions of open source.
- Pain point 1: медиа пишут «open source» — для compliance это разные legal buckets.
- Pain point 2: license ужесточилась vs K2 — больше не Modified MIT, custom doc с commercial thresholds.
- Pain point 3: 16.07–27.07 были только vendor claims — теперь community может audit weights и arch.
| Dimension | Open weight (Kimi K3) | Open source (OSI) |
|---|---|---|
| Model weights | Да (~1,56 TB) | Да |
| Technical report | Да | Да |
| Training data | Нет | Да |
| Full training code | Нет | Да |
| Commercial use | Да, с gates | License-dependent |
| Moonshot wording | «Open weight» | Не используется |
Fine-tune и self-host — да. Reproduce from scratch outside Moonshot — нет. Это operational delta, не semantic nitpick.
02 11-day timeline: API launch → full weight drop
От API-only до полного release: 11 дней. Для 2,8T MoE — агрессивный schedule.
- 16.07.2026 (WAIC eve): kimi.com, Kimi Work, Kimi Code, API — online inference only. Blog: «Kimi K3: Open Frontier Intelligence».
- 17.07: arch teardowns; Xinhua — largest open-weight model globally.
- 22–23.07: US-China distillation flare — Michael Kratsios (White House) accuses industrial distillation from Fable; Scott Bessent floats sanctions/Entity List.
- 27.07 (~23:00): full weights, technical report, MoonEP + AgentEnv (FlashKDA already open).
- 28.07: China MOFCOM pushback on «AI hegemonism»; domestic media on release specs.
| Spec | Value |
|---|---|
| Total params | 2,8 trillion (2,8T) |
| Active params | ~104 billion |
| MoE config | 896 routed experts, 16 active/token (+ shared experts) |
| Attention | KDA + Gated MLA hybrid |
| Context window | 1 000 000 tokens |
| Multimodal | Native vision (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights + MXFP8 activations (QAT from SFT) |
| Download size | ~1,56 TB (Hugging Face) |
Hard data: 2,8T / 104B active; MoE sparsity 1,8 % (16/896); first full 3T-class open-weight drop; HF trending #1 in 30 min.
03 Arch deep-dive: KDA, AttnRes, Per-Head Muon, Stable LatentMoE
K3 не просто scale-up — пересборка attention, residual path и optimizer.
Kimi Delta Attention (KDA)
Gated DeltaNet: scalar forget gate на весь memory state. KDA → channel-wise gating: per-dimension decay rate. Chunkwise DPLR recurrence — O(n) time, hardware-friendly. KDA layers interleaved с Gated MLA full-attention — enabler для 1M context при low KV cache footprint.
Attention Residuals (AttnRes)
Standard residuals uniformly accumulate layer outputs — early-layer signal dilutes с depth. AttnRes: selective, input-dependent aggregation across all prior layers. Overhead: RMSNorm + pseudo-query per layer (~negligible params). Moonshot claims ~25 % training efficiency gain for <2 % cost. Pseudo-queries init to zero → training start equivalent to uniform average.
Per-Head Muon
Muon optimizer extended per attention head — adaptive convergence paths. Pure training-time hack; invisible via API, но contributes к punch-above-active-param-count vs closed frontier models.
| Component | Function | Metric |
|---|---|---|
| Routing | 896 experts, 16 active/token | Sparsity 1,8 % |
| Quantile Balancing | Expert load без heuristic hyperparams | Stable MoE training |
| Per-Head Muon | Per-head optimizer | Scale training |
| Shared Experts | Baseline capacity | MoE robustness |
| Gated MLA | Selective global attention | Long-context hybrid |
04 MoonEP / FlashKDA / AgentEnv: infra release ≥ weights
Moonshot open-sourced не только checkpoint — training/serving stack, который делает 2,8T MoE feasible.
| Tech | Role | Key numbers |
|---|---|---|
| MoonEP | Comm lib для fine-grained MoE at scale | Temp duplicate overloaded experts; proven upper bound on redundant experts per rank |
| FlashKDA | CUTLASS KDA kernel (previously open) | H20 prefill 1,72×–2,22× vs flash-linear-attention; drop-in chunk_kda backend |
| AgentEnv | Firecracker microVM sandbox (w/ KVCache.ai) | Checkpoint 133 ms, resume 49 ms, memory overcommit до 6,5× (vendor claim, not independently verified) |
FlashKDA plugs into flash-linear-attention — zero code change для existing projects. AgentEnv targets massively parallel agentic RL с fast snapshot/fork/restore.
05 Benchmarks, license gates, cost-per-task math
Third-party evals first; vendor numbers labeled separately.
| Model | Score | Release |
|---|---|---|
| Claude Opus 5 | 97,0 % | 2026-07-24 |
| GPT-5.6 Sol | 96,2 % | 2026-07-09 |
| Claude Fable 5 | 95,0 % | 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): Fable 5 — 60; GPT-5.6 Sol — 59; K3 — ~57 (#3 overall, #1 open-weight); GLM-5.2 — 51; DeepSeek V4 Pro — 44. Cost/task: K3 ~$0,95 vs Fable ~$2,40 (−60 %) vs GLM-5.2 ~$0,47. Verdict: capability ceiling open-weight tier, not value pick. Arena.ai Frontend Code Arena: K3 #1.
| Gate | Threshold | Consequence |
|---|---|---|
| MaaS revenue | > $20M trailing 12 months (affiliates incl.) | Separate commercial agreement with Moonshot |
| Attribution / scale | > 100M MAU or > $20M monthly revenue | Prominent «Kimi K3» in product UI |
Для indie/small team — non-issue. Для MaaS competitors reselling K3 inference at Moonshot scale — clause #1 is the red line.
06 API pricing, 64-GPU floor, US-China context, 6-step deploy, FAQ
| Token type | Price |
|---|---|
| Input (cache hit) | $0,30 |
| Input (cache miss) | $3,00 |
| Output (incl. reasoning trace) | $15,00 |
Mooncake disaggregated serving: cache hit rate >90 % на typical coding workloads — effective input cost ближе к $0,30, не $3,00 headline. Self-host: Moonshot recommends 64+ accelerator cards supernode config; consumer/prosumer hardware — out of scope. OpenRouter: 7 providers, mostly list pricing.
US-China backdrop: release during WAIC 2026 window, days after distillation accusations (Kratsios/Bessent) and MOFCOM counter (28.07). Full weights + infra as engineering transparency play. +3 days: Alibaba Qwen3.8-Max-Preview (2,4T) — «3T club» race heating up.
6-step production integration:
- License audit: read custom doc; map MaaS gate + MAU threshold to business model.
- API key: platform.kimi.ai, fund account.
- OpenAI-compatible client:
base_url="https://api.moonshot.ai/v1", modelkimi-k3. - Cache optimization: repeat prompt prefixes for Mooncake hits; target >90 % cache hit.
- Stack benchmark: SWE-bench 93,4 % vs GLM-5.2 $0,47/task cost tradeoff.
- Hosting decision: API/OpenRouter for most; self-host only at 64+ GPU cluster; agent gateways on dedicated infra.
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 codebase..."}]
)Q: Kimi K3 open source?
A: Нет. Open weight — weights + report да, training data + full code нет. Moonshot never says open source.
Q: Commercial use free?
A: Да для большинства. Exceptions: MaaS >$20M/12mo или >100M MAU / >$20M monthly revenue.
Q: Self-host GPU count?
A: 64+ accelerators per Moonshot. Reality: API or OpenRouter.
Q: SWE-bench score?
A: 93,4 % (Vals AI, July 2026) — strongest open-weight, below Opus 5 and GPT-5.6 Sol.
Q: K3 vs K2 delta?
A: ~3× params vs K2.5; AttnRes + Per-Head Muon new; 1M context; MaaS license gate absent in K2 family.
API routing is fast, но prod agents hit hidden costs — shared VPS jitter, no launchd persistence, long-context RAM walls. Для 24/7 Kimi Code agents и MCP gateways: JEXCLOUD bare-metal Mac multi-region — dedicated Apple Silicon unified memory, no oversubscription, launchd-persistent gateways, 120s provisioning. Specs на странице pricing JEXCLOUD.