Kimi K3 open weight: 2.8T параметров, KDA и 93.4% SWE-bench — что Moonshot открыл 27 июля 2026

Около 22 мин чтения · MACCOME · Обновлено: 28 июля 2026

Для кого: инженеры, которые читают weight dumps, а не press release. Событие (27.07.2026, 23:00 Pekin): Moonshot AI выкатил Kimi K3 как open weight2.8T total params, ~104B activated, 1M token context, ~1.56TB weights на Hugging Face. Внутри: timeline API→weights за 11 дней, таблицы KDA/MoE/infra, independent benchmarks, license gates $20M, API tiers $0.30 / $3 / $15, порог 64 GPU supernode, 6-step runbook. Продолжение обзора от 16.07., distillation drama и OpenRouter июльского рейтинга.

insights

TL;DR — 30 секунд

  • Не OSI open source: Moonshot последовательно пишет open weight; training data и full training code — closed.
  • Spec sheet: 2.8T total, 896-of-16 MoE, 1M context, 1.56TB на HF — trending #1 за 30 минут после upload.
  • Benchmarks: SWE-bench Verified (Vals AI) 93.4%; Artificial Analysis Index ~57 — global #3, open weight #1.
  • License: MaaS revenue >$20M/12 mo или >100M MAU / >$20M monthly — extra clauses.
  • Deploy: официальный минимум 64 GPU supernode; для 99% команд — API или OpenRouter (7 providers).

Шесть типичных fail'ов после weight release K3

  1. Путать open weight с open source: медиа пишут «open source», Moonshot официально — только open weight: weights + tech report + infra, без training pipeline.
  2. Игнорировать license gates: MaaS >$20M/year или super-app scale (>100M MAU) — отдельные clauses; критично если вы конкурируете с Moonshot API.
  3. Верить vendor benchmarks: independent SWE-bench repro: K3 93.4%, не marketing ceiling; Claude Opus 5 держит 97%.
  4. Смотреть только capability, не $/task: K3 ~$0.95/task vs GLM 5.2 ~$0.47 — ceiling open-weight camp, не price/perf champ.
  5. Недооценивать self-host: 2.8T / 896 experts → Moonshot требует 64 GPU supernode minimum; consumer iron — nope.
  6. Agent gateway на ноутбуке: 1M context sessions требуют 24/7 uptime; sleep, jitter, scattered API keys — production killers, не model quality issues.

Timeline и core spec: 11 дней от API launch до full weights

  • 16.07.2026: K3 API launch на kimi.com, Kimi Code, Moonshot API — weights ещё closed.
  • 22–23.07.: US-side «industrial distillation» accusations; geopolitical backdrop перед WAIC 2026.
  • 27.07.2026, 23:00: full weights, tech report; MoonEP + AgentEnv open; FlashKDA уже был в open.
  • 28.07.2026: media/regulatory reactions; Hugging Face trending #1 за 30 мин (~1.56TB download).
ПараметрЗначение
Total params2.8T
Activated params~104B per token
ArchitectureMoE — 896 routing experts, 16 active (+ shared experts)
AttentionKimi Delta Attention (KDA) + Gated MLA
Context1M tokens
Weight formatMXFP4 weights + MXFP8 activations
Download size~1.56TB (Hugging Face)
LicenseCustom open-weight (не Modified MIT как K2)

Architecture stack: KDA, AttnRes, MoonEP

K3 переписывает три классических building block'а. Scalar forget gate в DeltaNet → per-channel KDA (DPLR recurrence, O(n) time, compressed KV cache). Attention Residuals (AttnRes) — selective aggregation предыдущих layer outputs вместо uniform residual stack: ~25% training efficiency при <2% overhead. Per-Head Muon — independent optimizer per attention head (training-only, API-invisible).

Stable LatentMoE: 896-of-16 routing, sparsity ~1.8%, shared experts для stability, Quantile Balancing + MoonEP для load balance на supernode scale.

ComponentRoleKey metric
KDA (FlashKDA)Linear-attention backend на CUTLASSPrefill на H20: 1.72–2.22x vs baseline
MoonEPFine-grained MoE all-to-all commsQuantile Balancing; theoretical bound redundant experts
AgentEnvFirecracker microVM sandbox (KVCache.ai)Checkpoint 133ms, restore 49ms (vendor claims)
FlashKDADrop-in backend для chunk_kdaЗамена flash-linear-attention baseline

Benchmark comparison и license gates

Priority: independent repro (Vals AI, Artificial Analysis). Vendor numbers — reference only.

ModelSWE-bench VerifiedTypeInput / Output (1M)
Claude Opus 597.0%Closed$5 / $25
GPT-5.6 Sol96.2%Closed
Claude Fable 595.0%Closed
Kimi K393.4%Open weight$3 / $15 (cache miss)
GLM 5.2Open weight$0.45 / $3.31
DeepSeek V476.2%Open weight~$0.05–0.14 / ~$0.24–0.28

License — два commercial gate: (1) Model-as-a-Service revenue >$20M за 12 месяцев — separate agreement с Moonshot. (2) >100M MAU или >$20M monthly revenue — visible «Kimi K3» attribution в product UI. Для SME и internal use — обычно friction-free.

API pricing и cost math

Token typePrice per 1MNote
Input (cache hit)$0.30Mooncake arch: >90% hit rate на code workloads
Input (cache miss)$3.00Full prefill
Output (incl. reasoning)$15.00Agent loops: output dominates bill

Endpoint: https://api.moonshot.ai/v1, model ID kimi-k3, OpenAI SDK compatible. Agent loop math (1M input + 3M output, cache miss): K3 ~$48; DeepSeek V4 Flash ~$0.90–1.50; Claude Opus 5 ~$80.

python
from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "system", "content": "Senior engineer. Type hints mandatory."},
        {"role": "user", "content": "Refactor this module for SWE-bench style tasks."},
    ],
    max_tokens=4096,
)
print(resp.choices[0].message.content)
warning

Prod note: self-host below 64 GPUs — not in Moonshot's design target. Если гоняете PII через API — document processing path, не только model capability slide.

6-step runbook: от weight dump до production agent stack

  1. License + use case audit: MaaS revenue, MAU, monthly revenue vs $20M gates; legal review если конкурируете с Moonshot API.
  2. Access path: official API (cache advantage), OpenRouter с fallback chain — см. OpenRouter API guide.
  3. Task-tier routing: bulk/code → DeepSeek V4 Flash; agent/coding → K3 max; critical → Claude Opus 5 — см. июльскую routing matrix.
  4. Benchmark your workloads: internal SWE subset или coding tasks; K3 vs GLM 5.2 cost curve — avoid over-provisioning.
  5. Deploy agent gateway: OpenClaw / Hermes / Kilo Code на dedicated host; API keys per env; 429/timeout fallback configured.
  6. 24/7 monitoring: daily budget, cache hit rate; host на MACCOME тарифах аренды.

Три цифры для tech review

  • Первый полностью открытый 3T-class model: 2.8T, 1.56TB, HF trending #1 за 30 минут.
  • 93.4% SWE-bench (Vals AI): +17.2pp vs DeepSeek V4 (76.2%); -3.6pp vs Opus 5 (97%).
  • Effective API price: при >90% cache hit input часто близок к $0.30 — output at $15 остаётся agent cost lever.

Self-host для большинства — fantasy; production agent stack всё равно требует always-on node. MACCOME Mac cloud: native macOS, SSH handoff, env isolation для stable K3 routing — цены и конфигурация.

Sources: Moonshot AI blog, Hugging Face (moonshotai), GitHub FlashKDA, Vals AI SWE-bench, Artificial Analysis. As of 28.07.2026.

FAQ

Kimi K3 — open source?

Нет в смысле OSI. Это open weight: weights, tech report и часть infra public; training data и full training code — closed.

Какие license gates на $20M?

MaaS revenue >$20M за 12 месяцев — separate Moonshot agreement. >100M MAU или >$20M monthly revenue — «Kimi K3» attribution в product UI.

Сколько GPU для self-host?

Moonshot рекомендует 64 GPU supernode minimum. Детали в обзоре от 16.07.

Что значит SWE-bench 93.4%?

Independent Vals AI repro на SWE-bench Verified — топ open weight, ниже frontier closed. Context: distillation debate.

K3 vs K2 — main delta?

K3 ~3x scale K2.5, AttnRes + Per-Head Muon added, 1M context; license с MaaS gates — новое vs K2 series.

Как крутить K3 agents 24/7?

Gateway на dedicated Mac cloud host; см. MACCOME цены аренды Mac mini для M4/M4 Pro нод.