On July 16, 2026, Moonshot AI quietly added "Kimi K3 is live" to the top of its API docs — no keynote, no fanfare — and shipped the world's largest open-source AI model at 2.8 trillion parameters. For AI developers and product leads choosing between frontier APIs, this article delivers: K3 specs and launch context, KDA / AttnRes / Stable LatentMoE architecture breakdown, full benchmarks vs Claude Fable 5 and GPT-5.6 Sol, pricing and four access paths, scenario selection matrix, July 27 full-weight open-source plan, plus a six-step getting-started checklist and FAQ.
Kimi K3 is Moonshot AI's 2.8 trillion (2.8T) parameter sparse Mixture-of-Experts (MoE) model — roughly 75% larger than the previous record holder DeepSeek V4 Pro (1.6T), 2.7x Xiaomi's open model (1.02T), and more than 7x Alibaba's (397B). At inference it activates 16 of 896 experts (1.8% sparsity), pairs that with a 1 million token context window and native vision, and targets complex coding, long-document reasoning, and knowledge work.
In one sentence: Kimi K3 is an open-source, vision-native, long-memory coding AI that costs 40% less than Claude Opus 4.8 — with full weights landing on Hugging Face July 27.
Size record streak: Kimi held the open-source parameter ceiling for 9 of the past 12 months
WAIC timing: Launch landed the night before WAIC 2026 (July 17–20) — a deliberate strategic signal
Commercial momentum: ARR passed $300M by June 2026; a sixth funding round closed at a $31.5B pre-money valuation
API-first revenue: API accounts for over 70% of revenue; overseas paid users grew 400%
Post-DeepSeek rebound: After losing significant market share over 18 months, K3 reasserts technical sovereignty — not just scale, but architecture
Full open-source pledge: Official WeChat announcement confirms complete weights on Hugging Face July 27
Kimi K3 is not raw parameter stacking. Engineering changes deliver roughly 2.5x better scaling efficiency versus Kimi K2.
Full attention makes KV cache memory grow quadratically on long contexts. KDA alternates linear and full attention layers in a 3:1 ratio: three cheap linear layers handle local structure; one full-attention layer preserves global information flow. Result: up to 75% less KV cache memory, up to 6.3x faster decoding at 1M tokens, and better performance than pure full-attention baselines across short, long, and RL-scaling regimes.
Standard residual connections spread information uniformly across depth, diluting early-layer representations in deeper layers. AttnRes adds selective retrieval — the model can pull high-value features from earlier layers directly across depth. That yields roughly 25% training efficiency gain 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 |
| SiTU (Sigmoid Tanh Unit) | Improved activation control |
| Gated MLA | Sharper attention selectivity |
Note: Benchmark figures below are self-reported by Moonshot AI. Each model used its own inference harness (K3 via Kimi Code, GPT via Codex, Claude via Claude Code). Independent third-party reproduction is still in progress.
| 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 (docs) | 91.1 | 89.8 | 85.8 | 87.9 | — |
Key takeaways: K3 leads SWE Marathon (sustained long coding) at 42.0 by a wide margin; edges Program Bench at 77.8; on FrontierSWE Fable 5 leads (86.6) while K3 still crushes GPT-5.6 Sol (71.3); OmniDocBench document understanding tops the field at 91.1. On Artificial Analysis Intelligence Index v4.1, K3 scores 57.1 — 4th overall, trailing Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9) by just 2.8 points from the top.
| 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 |
K3 matches Claude Sonnet 5 standard pricing ($3/$15) but delivers 5x the context. Cache hits drop to $0.30/M (one-tenth of list input price); coding workflows report 90%+ cache hit rates. China API: ¥20/M input, ¥100/M output, ¥2/M cache hit. Consumer kimi.com accounts are free with K3 at max reasoning effort; prepaid plans from ¥199 (promo through August 11).
Web / app (easiest): Visit kimi.com, sign up (Google supported). K3 runs at max reasoning effort by default
Get an API key: Create one at platform.kimi.ai
Configure an OpenAI-compatible client: Set base_url to https://api.moonshot.ai/v1, model to kimi-k3
OpenRouter route: Model ID moonshotai/kimi-k3 — official pricing, no markup, full 1M context
Enable cache optimization: Reuse system prompts and tool definitions in coding workflows; Mooncake disaggregated inference boosts cache hit rate
Mark July 27: Full weights on Hugging Face — evaluate private deployment (requires 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": "Analyze this code for me..."}]
)
| Scenario | Recommended Model | Why |
|---|---|---|
| Sustained long coding sessions | Kimi K3 | SWE Marathon leader; longest context |
| Complex repo-level bug fixing | Claude Fable 5 | FrontierSWE / SWE-bench Pro lead |
| Terminal / toolchain agents | GPT-5.6 Sol | Terminal Bench and Coding Agent Index lead |
| Ultra-long docs / multimodal | Kimi K3 | OmniDocBench #1; native vision + 1M context |
| Cost-sensitive workloads | DeepSeek V4 Pro | Output at $3.48/M — far below K3 |
| Self-hosted open weights (post 7/27) | Kimi K3 | Strongest downloadable open weights to date |
What it signals: China's open AI ecosystem is shifting from "cheap market share" to genuinely challenging the intelligence frontier — not vanity parameter counts, but architectural engineering.
If you plan to wire Kimi K3 into local agent pipelines or Kimi Code-style long coding sessions, running CLI agents on a laptop or unstable VPS often hits memory limits, dropped sessions, and missing Xcode/Metal toolchains; pure Linux VPS cannot cover iOS build and signing flows. For production environments that need stable SSH long sessions, DerivedData caching, and iOS CI/CD automation, NodeMini's Mac Mini cloud rental is usually the better fit — dedicated nodes, second-scale provisioning, and agents plus builds on one real Mac.
Yes. Register at kimi.com for free access to K3 at max reasoning effort. API usage is paid ($3/$15 per 1M tokens). For a stable agent runtime, see Mac Mini rental rates.
Full weights release on Hugging Face on July 27, 2026. Production deployment needs a supernode with 64+ accelerators; a laptop cannot host 2.8T parameters. Inference providers and well-resourced labs are the realistic self-host audience.
K3 has nearly 2x parameters (2.8T vs 1.6T), 1M vs 128K context, and stronger scores on several benchmarks. DeepSeek output costs only $3.48/M — much cheaper. Pick DeepSeek for cost; pick K3 for long code and document understanding.
Very useful for whole-repo analysis, long legal/research documents, and multi-turn agent memory. K3 bills at a flat rate with no length surcharge; KDA delivers up to 6.3x faster decoding at 1M tokens. More ops questions: help center.
Moonshot AI says low and high modes will ship in a future update. Only max mode is available today. Watch WAIC (July 17–20) and the July 27 weight release for updates.