2026 KIMI K3
2.8T_OPEN_
SOURCE_LLM.

Kimi K3 архитектура 2,8T open-source LLM и бенчмарки

Тезис: Три блокера при выборе closed-source API: достаточен ли контекст, держит ли coding-бенчмарк в production, когда доступны веса для self-host. 16.07.2026 Moonshot AI (月之暗面) тихо повесила «🎉 Kimi K3 is live!» в API-документации. Этот hardcore-разбор покрывает все опубликованные ключевые точки: спецификация 2,8T, KDA/AttnRes/Stable LatentMoE, полная benchmark-матрица, pricing, 5 путей интеграции, open weights 27.07, scenario matrix, FAQ.

Spec sheet · 30 сек

Release2026-07-16 · silent deploy · zero press event
Params2.8T · largest OSS model · +75% vs DeepSeek V4 Pro (1.6T)
Core1M ctx · native vision · MoE 896 experts / 16 active
APIInput $3/M · Output $15/M · cache hit $0.30/M
OSSFull weights 2026-07-27 on Hugging Face

1. Три engineering-вопроса — с цифрами

  1. Parameter sovereignty после DeepSeek. 18 месяцев потери доли рынка. K3 возвращает титул «largest OSS model» с 2.8T — релиз накануне WAIC 2026. Strategic signal: high.
  2. 1M ctx ≠ marketing number. Full attention → KV cache O(n²) memory blowup. KDA: -75% KV cache, 6.3× decode speedup at 1M tokens при фиксированном $3/M input.
  3. «Пишет код» ≠ «пишет код 4+ часа». SWE Marathon: K3 42.0 vs Fable 5 35.0 vs GPT-5.6 Sol 39.0. Ближе к real multi-hour agent sessions, чем one-shot SWE-bench.

2. Kimi K3 — формальное определение

16.07.2026, без keynote: tech blog + pricing page + model ID kimi-k3 — callable immediately.

Kimi K3 — крупнейший open-source LLM по числу параметров: 2.8 trillion (2.8T). +75% vs DeepSeek V4 Pro (1.6T), 2.7× Xiaomi (1.02T), 7×+ Alibaba (397B).

Sparse MoE: 896 experts, 16 active per forward pass (sparsity 1.8%). 1M token context window, native image/video understanding. Target: complex coding, long-doc reasoning, knowledge work.

One-liner: OSS heavyweight coding LLM с native vision и ultra-long memory — на 40% дешевле Claude Opus 4.8, full weights 27 июля.

3. Контекст релиза

  • Kimi series: 9 из 12 месяцев на вершине OSS scale leaderboard
  • Timing: eve of WAIC 2026 (Shanghai)
  • ARR (June 2026): >$300M · round 6 · valuation $31.5B
  • API revenue: >70% · international paid users +400%

Closed-source context: Claude Sonnet 5 / GPT-5.6 leak analysis, GPT-5.6 Sol Ultra math proof.

4. Architecture deep-dive

4.1 Kimi Delta Attention (KDA)

Standard Transformer full attention: compute ∝ seq_len². KDA — hybrid linear attention:

  • 3:1 ratio linear : full attention layers
  • KV cache memory: up to -75%
  • At 1M tokens: decode 6.3× faster
  • Beats full-attention baseline: short ctx, long ctx, RL scaling
# KDA layer stack (conceptual) # layers[0..2]: LinearAttention(K, V) # O(n) memory # layer[3]: FullAttention(Q, K, V) # O(n²) but sparse frequency # repeat 3:1 pattern across depth # KV cache footprint @ 1M ctx (Moonshot reported): # full_attn_only: ~100% baseline # KDA_hybrid: ~25% baseline (-75%)
Mental model: full attention = memorize every token. KDA = indexed retrieval + selective full recall on critical positions.

4.2 Attention Residuals (AttnRes)

Selective cross-depth retrieval — model pulls high-value representations from earlier layers without full re-computation. Reported: ~25% training efficiency gain, <2% compute overhead.

4.3 Stable LatentMoE

896 experts, 16 routed per token — 1.8% activation density:

ComponentFunction
Quantile BalancingExpert assignment from router score quantiles — zero heuristic hyperparams
Per-Head MuonPer attention-head optimizer — adaptive at scale
Sigmoid Tanh Unit (SiTU)Activation gating control
Gated MLAMulti-head latent attention selectivity

Net scaling efficiency vs Kimi K2: ~2.5×.

5. Benchmark matrix — raw numbers

BenchmarkKimi K3Claude Fable 5GPT-5.6 SolClaude Opus 4.8GLM-5.2
DeepSWE67.570.073.059.046.2
Program Bench77.876.877.671.963.7
Terminal Bench 2.188.384.688.884.682.7
FrontierSWE81.286.671.366.767.3
SWE Marathon42.035.039.040.013.0
BrowseComp91.288.090.484.3
Automation Bench30.829.129.727.212.9
GPQA-Diamond93.592.694.191.091.2
MMMU-Pro (Vision)81.681.283.078.9
OmniDocBench91.189.885.887.9

Parse notes:

  • SWE Marathon: K3 #1 (42.0) — best proxy for sustained coding sessions
  • Program Bench: K3 #1 (77.8) — marginal lead
  • FrontierSWE: Fable 5 leads (86.6); K3 (81.2) >> GPT-5.6 Sol (71.3)
  • OmniDocBench: K3 #1 (91.1) — vision + long ctx synergy
  • Artificial Analysis Intelligence Index v4.1: K3 57.1 (rank 4) — behind Fable 5 (59.9), GPT-5.6 Sol (58.9)

Caveat: Moonshot self-reported. Different inference harnesses (K3: Kimi Code, GPT: Codex, Claude: Claude Code). Independent reproduction pending.

6. Pricing — $/M token breakdown

ModelInput ($/M)Output ($/M)Cache hitContext
Kimi K3$3.00$15.00$0.301M
Claude Sonnet 5$3.00 (promo $2)$15.00 (promo $10)200K
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 standard ($3/$15), 5× context window
  • Cache hit $0.30/M — coding scenarios >90% hit rate
  • China API: ¥20/M input · ¥100/M output · cache ¥2/M
  • kimi.com free tier; prepaid from ¥199 (promo until Aug 11)

7. Integration paths — 5 routes

  1. Kimi Web/App: kimi.com — Google OAuth, max reasoning default.
  2. Official API: key at platform.kimi.ai, OpenAI-compatible:
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 diff..."}], max_tokens=8192, # 1M ctx: pass full repo as system context # cache hit $0.30/M on repeated prefixes )
  1. OpenRouter: moonshotai/kimi-k3 — official pricing, full 1M ctx.
  2. Cursor multi-route: K3 primary for long-ctx coding; complex repo bugs → Claude Fable 5 fallback.
  3. Post 27.07: self-host — Hugging Face weights (64+ GPU supernode for production).

8. Scenario routing matrix

WorkloadRoute toRationale
Sustained long codingKimi K3SWE Marathon #1 · longest ctx
Complex repo-level bugfixClaude Fable 5FrontierSWE +5.4pt lead
Terminal/toolchain agentsGPT-5.6 SolTerminal Bench 2.1 leads
Ultra-long docs / multimodalKimi K3OmniDocBench #1 · native vision + 1M
Cost-sensitive batchDeepSeek V4 ProOutput $3.48/M vs $15/M
OSS self-host (near-term)Kimi K3 (post 27.07)Strongest downloadable weights

9. Open weights — 27 July milestone

Moonshot official commitment: 2026-07-27 full model weights on Hugging Face. Post-release K3 becomes:

  • Largest downloadable OSS model
  • First OSS weights above 2T params
  • New community fine-tuning baseline

Quantization-aware training: MXFP4 weights, MXFP8 activations. HF variants: MXFP4/NVFP4. Expected day-1: vLLM, SGLang.

# Expected self-host stack (post 27.07) # vLLM launch example (speculative): python -m vllm.entrypoints.openai.api_server \ --model moonshotai/Kimi-K3 \ --tensor-parallel-size 64 \ --max-model-len 1048576 \ --quantization mxfp4 # MLX community quant path (Mac): # watch: huggingface.co/mlx-community for GGUF/Q4_K_M builds

Timeline: Jul 17–20 (WAIC) → Jul 27 (weights drop).

10. FAQ — technical reference

Free tier?
kimi.com free account. API: $3/$15 per 1M tokens.

Local deploy?
Weights Jul 27. Production: 64+ accelerators — not laptop territory.

vs DeepSeek V4 Pro?
K3: 2× params, 1M vs 128K ctx, stronger benchmarks. DeepSeek: $3.48/M output — 4.3× cheaper.

1M ctx useful?
Full codebase ingest, legal/research docs, multi-turn agent memory — flat pricing, no length surcharge.

Low/high reasoning?
Moonshot announced for future updates. Max only today.

11. OSS ecosystem signal

Kimi K3 ≠ parameter vanity project. KDA, AttnRes, Stable LatentMoE are real engineering deltas. On sustained coding, doc understanding, and reasoning — K3 competes with closed-source flagships at reasonable API pricing with full OSS commitment.

Post Jul 27: first 2T+ downloadable baseline for enterprise on-prem, vertical fine-tuning, API-independent agent stacks. Mac devs: K3 API in Cursor for long-ctx coding is the pragmatic path now; track MLX/llama.cpp quant adapters post-weight-drop — see DeepSeek custom chip & inference infra.

12. Verdict: API anywhere, agent stress on remote Mac

kimi.com signup, API key, OpenRouter routing — Windows/Linux sufficient. For Cursor + K3 1M ctx, Xcode agent integration, MLX quant validation — Apple Silicon unified memory + Metal remains lowest-friction path.

Optimal split: local API dev, Kimi Code stress tests, multi-repo SWE Marathon regression, 1M doc batch on MACGPU remote Mac mini M4 — on-demand SSH, agent load isolated from primary machine.