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 weight — 2.8T 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.
TL;DR — 30 секунд
| Параметр | Значение |
|---|---|
| Total params | 2.8T |
| Activated params | ~104B per token |
| Architecture | MoE — 896 routing experts, 16 active (+ shared experts) |
| Attention | Kimi Delta Attention (KDA) + Gated MLA |
| Context | 1M tokens |
| Weight format | MXFP4 weights + MXFP8 activations |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom open-weight (не Modified MIT как K2) |
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.
| Component | Role | Key metric |
|---|---|---|
| KDA (FlashKDA) | Linear-attention backend на CUTLASS | Prefill на H20: 1.72–2.22x vs baseline |
| MoonEP | Fine-grained MoE all-to-all comms | Quantile Balancing; theoretical bound redundant experts |
| AgentEnv | Firecracker microVM sandbox (KVCache.ai) | Checkpoint 133ms, restore 49ms (vendor claims) |
| FlashKDA | Drop-in backend для chunk_kda | Замена flash-linear-attention baseline |
Priority: independent repro (Vals AI, Artificial Analysis). Vendor numbers — reference only.
| Model | SWE-bench Verified | Type | Input / Output (1M) |
|---|---|---|---|
| Claude Opus 5 | 97.0% | Closed | $5 / $25 |
| GPT-5.6 Sol | 96.2% | Closed | — |
| Claude Fable 5 | 95.0% | Closed | — |
| Kimi K3 | 93.4% | Open weight | $3 / $15 (cache miss) |
| GLM 5.2 | — | Open weight | $0.45 / $3.31 |
| DeepSeek V4 | 76.2% | 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.
| Token type | Price per 1M | Note |
|---|---|---|
| Input (cache hit) | $0.30 | Mooncake arch: >90% hit rate на code workloads |
| Input (cache miss) | $3.00 | Full prefill |
| Output (incl. reasoning) | $15.00 | Agent 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.
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)
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.
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.