Главная / Блог / Kimi K3
ENGINEERING BLOG · 2026.07.17

Kimi K3:
обзор открытой модели на 2,8 трлн параметров

Ночью 16 июля 2026 Moonshot AI включила Kimi K3 без keynote — баннер в API-доке, tech blog и рабочий endpoint. Этот обзор для ML-инженеров, SRE и agent-разработчиков, которым нужны не маркетинговые тезисы, а цифры: 2,8 трлн параметров, механизм Kimi Delta Attention, сравнение с Claude Fable 5 / GPT-5.6 Sol, цены API, пошаговая интеграция и дата open weights — 27 июля 2026.

01

Kimi K3 — крупнейшая open-source LLM на сегодня: 2,8T параметров (+75 % к DeepSeek V4 Pro 1,6T). Sparse MoE: 896 experts, на forward pass активируются 16 (sparsity 1,8 %). Контекст: 1 048 576 tokens. Native vision (image/video). Фокус — long-horizon coding и knowledge work.

Kimi K3 — ключевые параметры (16.07.2026)
Параметр Значение Контекст
Total params2,8TRecord open source
MoE config896 / 16 activeExpert routing 1,8 %
Context window1 048 5765× Claude Sonnet 5
MultimodalVision nativeOmniDocBench 91,1
Model IDkimi-k3OpenAI-compatible API
Open weights27.07.2026Hugging Face — full checkpoint

Стратегический контекст: Kimi лидировала по размеру open models 9 из 12 месяцев; релиз — накануне WAIC 2026. ARR >$300M (июнь 2026), valuation $31,5B pre-money, API revenue >70 %, international paid users +400 %.

Production pain points, которые specs не снимают:

  • Cloud lock-in до 27.07.: pipeline зависит от Moonshot infra — latency, quotas, pricing drift.
  • 1M context ≠ zero engineering cost: без persistent dev environment (bare-metal Mac, стабильный toolchain) agent resume и tool integration ломаются.
  • Self-host barrier: Moonshot указывает 64+ accelerators для production inference — consumer hardware out of scope.
  • Benchmark harness mismatch: Kimi Code vs Claude Code vs Codex — независимые reproductions in progress.
  • Metal/Xcode stack: API-модель не заменяет native Apple Silicon для iOS CI/CD и long-running agents на физическом Mac.

02

K3 — не naive scaling. Сравните с моделями из OpenRouter June ranking: Moonshot оптимизирует long-context decode и training stability при extreme sparsity.

Kimi Delta Attention (KDA)

Full attention — O(n²); при 1M tokens KV-cache съедает HBM. KDA чередует linear и full attention в ratio 3:1:

  • 3 linear layers — cheap local structure; 1 full layer — global information flow;
  • KV-cache memory: до −75 %;
  • Decode at 1M tokens: до 6,3× быстрее;
  • Beat full-attention baseline на short/long context и RL scaling.

Attention Residuals (AttnRes)

  • Standard residual path размывает early-layer representations на глубине;
  • AttnRes — selective retrieval high-value features через depth skip;
  • Training efficiency +25 %, compute overhead <2 %.

Stable LatentMoE

Stable LatentMoE — supporting techniques
Technique Function
Quantile BalancingExpert assignment из router quantiles — без heuristic hyperparams
Per-Head MuonPer-head optimizer для stable large-scale training
Sigmoid Tanh Unit (SiTU)Activation control
Gated MLAAttention selectivity
Net effect~2,5× scaling efficiency vs Kimi K2

03

Self-report Moonshot, 16.07.2026. Для repo-level bugfix — Claude Code на Mac mini.

Core benchmarks (Moonshot self-report)
Benchmark Kimi K3 Claude Fable 5 GPT-5.6 Sol Claude Opus 4.8
DeepSWE67,570,073,059,0
Program Bench77,876,877,671,9
Terminal Bench 2.188,384,688,884,6
FrontierSWE81,286,671,366,7
SWE Marathon42,035,039,040,0
BrowseComp91,288,090,484,3
Automation Bench30,829,129,727,2
GPQA-Diamond93,592,694,191,0
MMMU-Pro81,681,283,078,9
OmniDocBench91,189,885,887,9

Geek read: K3 лидирует SWE Marathon (42,0) — multi-hour coding sessions. Program Bench 77,8 — #1. FrontierSWE — Fable 5 territory (86,6). AA Intelligence Index v4.1: K3 57,1, rank #4.

Self-report, разные harnesses — ждите independent reproduction перед production cutover.

04

Price/context matrix (USD per 1M tokens)
Model Input Output Cache hit Context
Kimi K3$3,00$15,00$0,301M
Claude Sonnet 5$3,00$15,00200K
Claude Opus 4.8$5,00$25,00200K
GPT-5.5$5,00$30,00400K
DeepSeek V4 Pro$1,74$3,48$0,145128K
Kimi K2.6$0,95$4,00$0,16256K

K3 = Sonnet 5 pricing ($3/$15), но 5× context. Cache hits от $0,30/M; Moonshot claims 90 %+ cache hit в coding. China: ¥20/¥100/¥2 per M. Consumer kimi.com — free tier.

6 шагов интеграции:

  1. Регистрация Moonshot: platform.kimi.ai или kimi.com.
  2. API key: generate в dashboard, store securely (one-time display).
  3. Model ID: kimi-k3; OpenRouter: moonshotai/kimi-k3.
  4. OpenAI-compatible client: base_url = https://api.moonshot.ai/v1.
  5. Smoke test: short prompt — latency и quota check.
  6. Cache-aware workflow: reuse system prompts и tool definitions для 90 %+ cache hits.
kimi_k3_api.py
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": "Разбери этот код..."}]
)
print(response.choices[0].message.content)

05

Model selection by workload
Scenario Pick Why
Long-horizon codingKimi K3SWE Marathon #1, max context
Repo-level bugfixClaude Fable 5FrontierSWE lead
Terminal/tool agentsGPT-5.6 SolTerminal Bench 2.1
Long docs + multimodalKimi K3OmniDocBench 91,1
Cost-sensitiveDeepSeek V4 ProOutput $3,48/M
Self-host post-27.07.Kimi K3Strongest open weights to date

Moonshot: full weights on Hugging Face 27 июля 2026 — first open model >2T. Training: MXFP4 weights + MXFP8 activations. Expect vLLM/SGLang support и MXFP4/NVFP4 quants. Timeline: WAIC 17–20.07 → 27.07. open weights.

06

  • Total params: 2,8T
  • MoE: 896 experts, 16 active (1,8 % sparsity)
  • Context: 1 048 576 tokens
  • KDA: KV −75 %, decode +6,3× at 1M
  • AttnRes: training +25 %, overhead <2 %
  • Scaling vs K2: ~2,5×
  • API: $3 / $15 / $0,30 per M tokens
  • China: ¥20 / ¥100 / ¥2
  • ARR Jun 2026: >$300M; valuation $31,5B
  • AA Index v4.1: 57,1 (#4)
  • Open weights: 27.07.2026 (Hugging Face)

Re-verify links after upstream updates:

Official:

Moonshot AI — Kimi K3 tech blog

Kimi API Platform — docs and pricing

Third-party:

Artificial Analysis — Intelligence Index v4.1

OpenRouter — moonshotai/kimi-k3

Kimi K3 поднимает планку open source — но pure API = latency, quotas, fragile toolchain; self-host до 27.07. impossible, после — supernode GPU farm; shared cloud VM режет Metal performance и ломает long-running agents. Для native Apple Silicon, 7×24 uptime, stable iOS CI/CD (Kimi Code / Claude Code на persistent physical Mac) — ZUKCLOUD bare-metal Mac mini cloud nodes: dedicated hardware, zero hypervisor tax, flexible billing. Pricing, Order, bare-metal manifesto.

07

Q: Kimi K3 бесплатен?
A: kimi.com — free consumer tier (max reasoning). API: $3/$15 per M tokens.

Q: Local deploy?
A: Weights с 27.07.2026 на Hugging Face. Production: 64+ accelerators.

Q: K3 vs DeepSeek V4 Pro?
A: K3 — 2× params, 5× context, stronger coding benches. DeepSeek — output $3,48/M. См. DeepSeek chip ecosystem.

Q: Зачем 1M context?
A: Whole-repo analysis, long legal/scientific docs, agent memory. Flat pricing + high cache rate.

Q: Когда low/high reasoning?
A: Moonshot анонсировала low и high для future updates; сейчас только max.