2,8T paramètres · KDA · 1M tokens · Benchmarks · Poids ouverts 27/07
Dans la nuit du 16 juillet 2026, Moonshot AI a activé Kimi K3 discrètement — sans keynote, mais avec 2,8 billions de paramètres, le plus grand modèle open source jamais publié. Pour choisir entre Kimi K3, Claude Fable 5 et GPT-5.6 Sol, ou attendre les poids du 27 juillet : architecture KDA/AttnRes/Stable LatentMoE, tableaux de benchmarks, matrices tarifaires, quatre voies d'intégration et runbook en six étapes. SWE Marathon, OmniDocBench, tarifs Chine et signal WAIC inclus.Données au 17.07.2026
On July 16, 2026, Moonshot AI quietly shipped Kimi K3: a technical blog, a pricing page, and a callable model ID kimi-k3. The low-key launch contrasts sharply with its 2.8T parameter count — roughly 75% larger than prior record holder DeepSeek V4 Pro (1.6T), 2.7x Xiaomi's open model (1.02T), and more than 7x Alibaba's (397B).
| Spec | Details |
|---|---|
| Total parameters | 2.8 trillion (2.8T) |
| Architecture | Kimi Delta Attention + Attention Residuals + Stable LatentMoE |
| Active experts | 16 of 896 per forward pass (1.8% sparsity) |
| Context window | 1,048,576 tokens (1M — roughly five full novels) |
| Input modalities | Text, image, video (native vision) |
| Reasoning modes | Max effort only today; low/high coming later |
| API pricing | $3 / $15 per 1M tokens (input/output) |
| Open weights | July 27, 2026 — full release on Hugging Face |
Bottom line: Kimi K3 is an open-weight, vision-native, long-memory coding model priced roughly 40% below Claude Opus 4.8, with full weights promised on July 27.
The launch context matters too. DeepSeek's rise squeezed Moonshot over the past 18 months, but K3 is a sharp counterpunch — Kimi held the open-source size crown for 9 of the past 12 months; timing landed the night before WAIC 2026; ARR crossed $300M by June 2026; a sixth funding round this year valued the company at $31.5B pre-money; API revenue exceeds 70% of total; overseas paid users grew 400%.
Before you pick a model, developers usually hit five pain points:
Context too short: 200K windows truncate large repo analysis; long coding tasks need constant context stitching.
Closed API costs spiral: Claude Opus 4.8 output runs $25/M; high-frequency agents can rack up five-figure monthly bills.
Open models lagged: Even DeepSeek V4 Pro — previously the largest open model — trailed closed flagships on long tasks like SWE Marathon.
Self-reported benchmarks: Each vendor uses a different harness (Kimi Code / Codex / Claude Code) — cross-reading required.
Self-hosting barrier: A 2.8T model needs a 64+ accelerator supernode; until July 27, API or kimi.com only.
Kimi K3 is not raw parameter stacking — it introduces engineering-level changes across attention, residual connections, and MoE routing, delivering roughly 2.5x better scaling efficiency versus Kimi K2.
Full attention makes KV cache memory grow quadratically on long contexts. KDA is a hybrid linear attention mechanism alternating linear and full-attention layers at a 3:1 ratio — three linear layers handle local structure, one full-attention layer preserves global information flow. Results: up to 75% less KV cache memory; up to 6.3x faster decoding at 1M tokens; beats pure full-attention baselines on short, long, and RL scaling workloads.
Standard residual connections accumulate uniformly by depth, diluting high-value representations from early layers. AttnRes adds selective retrieval — the model can pull valuable early-layer features across depth directly, yielding roughly 25% training efficiency gains with under 2% extra compute.
| Technique | Role |
|---|---|
| Quantile Balancing | Derives expert allocation from router score quantiles — no heuristic hyperparameters |
| Per-Head Muon | Per-attention-head optimization for more adaptive large-scale training |
| Sigmoid Tanh Unit (SiTU) | Improved activation function control |
| Gated MLA | Sharper attention selectivity |
If full attention is memorizing every conversation detail at once, KDA is an efficient assistant — fast indexing most of the time, precise recall when it counts. That is how K3 delivers a 1M token window at mainstream pricing.
Core benchmarks below are Moonshot self-reported (K3 via Kimi Code, GPT via Codex, Claude via Claude Code). Independent third-party reproduction is still in progress — treat these as directional.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | GLM-5.2 |
|---|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | 46.2 |
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 | 63.7 |
| Terminal Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 | 82.7 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 67.3 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 13.0 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | — |
| Automation Bench | 30.8 | 29.1 | 29.7 | 27.2 | 12.9 |
| GPQA-Diamond | 93.5 | 92.6 | 94.1 | 91.0 | 91.2 |
| MMMU-Pro (vision) | 81.6 | 81.2 | 83.0 | 78.9 | — |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 | — |
Note: Vendor-reported numbers using different harnesses are not directly comparable. On deep reasoning tasks like HLE-Full, Claude Fable 5 still leads by a wide margin (53.3 vs K3's 43.5).
K3 standard pricing matches Claude Sonnet 5 ($3/$15) but ships 5x the context; cache hits drop to $0.30/M, and coding workflows report cache hit rates above 90% — making effective input cost very low.
| Model | Input ($/M) | Output ($/M) | Cache hit input | Context |
|---|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 | 1M |
| Claude Sonnet 5 | $3.00 (promo $2) | $15.00 (promo $10) | — | 200K |
| Claude Opus 4.8 | $5.00 | $25.00 | — | 200K |
| GPT-5.5 | $5.00 | $30.00 | — | 400K |
| DeepSeek V4 Pro | $1.74 | $3.48 | $0.145 | 128K |
| Kimi K2.6 | $0.95 | $4.00 | $0.16 | 256K |
China domestic API pricing: input ¥20/M, output ¥100/M, cache hit ¥2/M. Consumer kimi.com is free with a basic account; prepaid plans start at ¥199 (promo through August 11).
| Scenario | Recommended model | Why |
|---|---|---|
| Sustained long-code tasks | Kimi K3 | SWE Marathon leader, longest context |
| Complex repo-level bug fixing | Claude Fable 5 | FrontierSWE / SWE-bench Pro lead by a wide margin |
| Terminal/toolchain-heavy agents | GPT-5.6 Sol | Terminal Bench and Coding Agent Index lead |
| Ultra-long docs / multimodal understanding | Kimi K3 | OmniDocBench leader, native vision + 1M context |
| Cost-sensitive workloads | DeepSeek V4 Pro | Output only $3.48/M |
| Open self-hosting (post July 27) | Kimi K3 | Strongest open weights available |
K3 is live on kimi.com, the Kimi app, Kimi Code, and the Moonshot API. Here is a six-step runbook to go from zero to production:
Try it in browser or app: Visit kimi.com, sign up (Google login supported). K3 runs at max reasoning by default — no credit card required.
Get an API key: Create a project at platform.kimi.ai, copy your key, and confirm balance and rate limits.
Configure the OpenAI-compatible SDK: Set base_url=https://api.moonshot.ai/v1 and model ID kimi-k3 (see code below).
OpenRouter (optional): Model ID moonshotai/kimi-k3 at official $3/$15 pricing, no markup, full 1M context.
Optimize cache hit rate: Keep system prompts stable in coding workflows; leverage Mooncake disaggregated inference. Cache hits cost only $0.30/M input — Moonshot reports 90%+ hit rates in coding scenarios.
Mark the July 27 milestone: Full weights land on Hugging Face; expect day-one support in vLLM, SGLang, and transformers. Production deployment needs a 64+ accelerator supernode.
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": "Help me analyze this code..."}]
)
Moonshot's official WeChat announcement confirms full model weights on July 27. K3 will become the largest downloadable open model ever, the first open weights above 2T parameters, and a new fine-tuning base for the open community. Weights trained in MXFP4 with MXFP8 activations — quantization-friendly, with MXFP4/NVFP4 variants expected on Hugging Face.
Kimi K3 is not a vanity parameter count — it ships real engineering innovation, matches or beats some closed flagships on long-code and document understanding, and commits to full open weights. That signals China's open AI ecosystem moving from "cheap market share" to genuinely challenging the frontier.
L'API couvre l'inférence K3, mais chaînes de signature iOS, builds Xcode locaux, Metal et CI 24/7 exigent du macOS physique. Les VM ajoutent overhead et risque EULA. Pour une CI/CD iOS et des agents stables, la location cloud Mac Mini VpsMesh est généralement le meilleur choix — Apple Silicon bare-metal, root, coût mensuel prévisible.
Sources: Moonshot AI official blog · Kimi API Platform docs · MarkTechPost · VentureBeat · SCMP · Artificial Analysis · OpenRouter pricing page. Benchmarks are Moonshot self-reported as of July 16, 2026.
Oui — compte gratuit kimi.com avec raisonnement max. Environnement build : tarifs Mac Mini M4.
Full weights arrive on July 27, 2026 via Hugging Face. Production inference needs a 64+ accelerator supernode — not a laptop. Until then, use the API or kimi.com.
K3 has nearly 2x parameters (2.8T vs 1.6T), 1M vs 128K context, and stronger benchmarks. DeepSeek output is only $3.48/M — still the pick for cost-sensitive workloads. See our OpenRouter model selection guide.
Yes — especially for whole-repo analysis, long legal or research documents, and multi-turn agents that need persistent memory. K3 charges flat per-token rates with no length surcharge, so using the full window is practical.
Moonshot says low and high effort modes are coming in a future update. Only max is available today. For a 24/7 build environment, see Mac Mini M4 rental pricing.
ID modèle moonshotai/kimi-k3. Détails dans le centre d'aide.