Главная / Блог / K3 dump
ENGINEERING_BLOG · 2026.07.28

Kimi K3: полный дамп 2,8T MoE, миллион токенов за 11 дней после API

KIMI K3 · TOTAL PARAMS
2.8T

Первый полностью открытый 3T-class чекпоинт — ~1,56 TB на Hugging Face.

27 июля 2026, ~23:00 по Пекину: Moonshot AI выкатила full weights Kimi K3, tech report и одновременно open-sourced MoonEP, FlashKDA, AgentEnv — инфраструктуру, на которой тренировали MoE с 2,8 трлн total params (~104B active), 1M token context window и native vision. От API launch (16 июля) до полного дампа — 11 дней. Статья для ML/dev-команды: timeline, KDA/AttnRes/Per-Head Muon stack, SWE-bench 93.4%, два commercial threshold в custom license, 64-GPU self-host reality check. TL;DR: K3 — потолок open-weight capability, но не OSI open-source; для 99% команд API/OpenRouter beats 1,56 TB + 64 GPU cluster.

SECTION 01 5 мифов, которые сломают ваш K3 rollout

  • open weight ≠ open source: Moonshot последовательно пишет «open weight». Weights + report + partial infra — да. Training corpus + full code — нет. OSI не выполнен.
  • Modified MIT мёртв: K3 на custom license. Пороги: MaaS revenue >$20M/year или MAU >100M/monthly revenue >$20M.
  • Self-host — не для mortals: 896 experts, 1,56 TB weights, official rec: 64+ GPU supernode. Ваш homelab не потянет.
  • Bang/buck: ~$0.95/task (AA) — на 60% дешевле Claude Fable 5, но дороже GLM-5.2 (~$0.47). Прогноз от 22 июля про Modified MIT устарел.
  • Geo context: Distillation drama 22–23 июля + ответ MOFCOM 28 июля — релиз weights как political signal, не только tech drop.

SECTION 02 Timeline: API → full dump за 11 дней

  • 16.07 (pre-WAIC): K3 live на kimi.com, Kimi Work/Code, API — weights closed.
  • 17.07: Industry teardown; Xinhua: «крупнейший open-weight model».
  • 22–23.07: US-CN distillation spat эскалирует.
  • 27.07 ~23:00: Full weights + report; MoonEP + AgentEnv open (FlashKDA уже был).
  • 28.07: MOFCOM response; media follow-up. +3 дня: Alibaba Qwen3.8-Max-Preview (24T).
Kimi K3 — core specs
Параметр Значение
Total params 2,8 трлн (2.8T)
Active params ~104 млрд
Architecture MoE (Mixture-of-Experts)
Expert config 896 routed experts, 16 active/token (+ shared)
Attention KDA + Gated MLA
Context 1 048 576 tokens
Multimodal Native vision (ViT-V2, 27 layers)
Weight format MXFP4 weights + MXFP8 activations (QAT from SFT)
Weight size ~1,56 TB (Hugging Face)
License Custom (open weight, not open source)

SECTION 03 KDA / AttnRes / MoonEP — что под капотом

K3 не просто накрутил params — переписал attention, residual connections и optimizer. Это и даёт gap в SWE-bench и AA Index.

  • KDA (Kimi Delta Attention): Per-channel decay вместо scalar forget gate; chunkwise DPLR для linear-time recurrence; interleave с Gated MLA — 1M context при сжатом KV cache.
  • AttnRes: Selective aggregation предыдущих layer outputs по input; +25% training efficiency, <2% overhead (RMSNorm + pseudo-query per layer).
  • Per-Head Muon: Independent optimizer per attention head — pure training trick, inference-transparent.
  • Stable LatentMoE: 896→16 routing (~1.8% density), Quantile Balancing + MoonEP с proven upper bound redundant experts/node.
Open infra drop (27.07.2026)
Stack Role Numbers
MoonEP Fine-grained MoE comms lib Temp expert replication; even token distribution; redundant expert bound proof
FlashKDA CUTLASS KDA kernels H20 prefill 1.72–2.22× vs flash-linear-attention baseline; drop-in chunk_kda backend
AgentEnv Agent sandbox w/ KVCache.ai (Firecracker microVM) Parallel Agent RL; checkpoint 133ms, recovery 49ms, memory oversub 6.5×
SWE-bench Verified — independent rerun (Vals AI, July 2026)
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): Fable 5 = 60, GPT-5.6 Sol = 59, K3 ≈ 57 (#3 global, #1 open-weight), GLM-5.2 = 51. Arena.ai Frontend Code Arena — K3 на #1.

K3 = open-weight ceiling, не price/perf king. Нужен frontier-near coding — берите K3. Нужен bulk на дешёвом tier — GLM-5.2.

SECTION 04 6-step integration: API, license, production

  1. Parse custom license: MaaS >$20M/year или MAU >100M/monthly revenue >$20M — legal review mandatory.
  2. Grab Moonshot API key: Model ID kimi-k3, endpoint https://api.moonshot.ai/v1.
  3. Wire OpenAI-compatible client: Swap base_url + api_key; Chat Completions logic unchanged.
  4. Enable Prompt Caching: Mooncake disaggregated inference — >90% cache hit на coding workloads → effective ~$0.30/M input.
  5. Route via OpenRouter: Multi-model fallback + unified billing — см. OpenRouter guide (7 providers уже hostят K3).
  6. Self-host TCO math: 64 GPU + 1.56 TB + power + SRE vs API spend. No supernode → stick to API.
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": "Explain Kimi Delta Attention"}]
)
print(response.choices[0].message.content)

SECTION 05 Pricing, hard numbers, primary sources

Kimi K3 API pricing (per 1M tokens)
Token type Price
Input (cache hit) $0.30
Input (cache miss) $3.00
Output (incl. reasoning) $15.00
  • Weights: 1.56 TB on HF; trending #1 within 30 min post-upload.
  • Sparsity: 16/896 experts active — 1.8% density.
  • FlashKDA: 1.72–2.22× prefill on H20.
  • Task cost: ~$0.95 (AA) vs Fable 5 ~$2.4 (−60%) vs GLM-5.2 ~$0.47.
  • Competition: Qwen3.8-Max-Preview (24T) три дня спустя — «3T club» heating up.

Primary sources — re-verify после публикации:

Moonshot AI — Kimi K3 tech blog

Hugging Face moonshotai (K3 weights)

GitHub moonshotai/FlashKDA

Artificial Analysis Intelligence Index

Production pattern: K3 inference через Moonshot API/OpenRouter для 1M context, agent orchestration + iOS CI/CD на VPSNIX M4/M4 Pro physical nodes — 100% Apple hardware, full root, zero hypervisor tax, elastic day/month billing. Cloud LLM не подпишет ваш iOS binary и не отдебажит Metal на virtualized macOS. K3 wants 64 GPUs; macOS agents want bare-metal Apple silicon. Pricing.

SECTION 06 FAQ

K3 — open-source?

Нет в OSI-смысле. Open weight: weights, report, partial infra public; training data + full code — нет.

Commercial use free?

Most cases — да. MaaS >$20M/year или MAU >100M/monthly revenue >$20M — extra license terms.

GPU count для self-host?

Official: 64+ GPU supernode. Majority: API или OpenRouter.

K3 strong?

Strongest open-weight, AA Index ~57 (#3 global). Дороже GLM-5.2 per task.

K3 vs K2?

~3× params vs K2.5, AttnRes, Per-Head Muon, 1M context, vision. License: new MaaS thresholds.