On the night of July 16, 2026, Moonshot AI posted a banner at the top of its API documentation — "🎉 Kimi K3 is live!" — with no keynote, no hype cycle, just a technical blog post, a pricing page, and an immediately callable model ID kimi-k3. The quiet launch contrasts sharply with a 2.8-trillion-parameter footprint.

One-line definition: Kimi K3 is currently the largest open-source AI model by parameter count — 2.8T parameters, roughly 75% larger than DeepSeek V4 Pro (1.6T), 2.7x Xiaomi's open model (1.02T), and more than 7x Alibaba (397B). It uses sparse MoE architecture, activating 16 of 896 experts at inference; paired with a 1-million-token context window (roughly five full copies of Dream of the Red Chamber) and native vision understanding, it targets complex coding, long-document reasoning, and knowledge work. Full weights open on July 27, at pricing roughly 40% below Claude Opus 4.8.

Why this launch matters: Moonshot AI spent the past 18 months under pressure from DeepSeek's rise. K3 is a sharp counterpunch:

Kimi K3 is not simple parameter stacking — three engineering innovations address real bottlenecks in long context and ultra-sparse MoE training.

2.1 Kimi Delta Attention (KDA) — Hybrid Linear Attention

Standard full attention makes KV-cache memory grow quadratically with context — catastrophic at 1M tokens. KDA alternates linear and full attention layers in a 3:1 ratio: three linear layers handle local structure cheaply; one full attention layer preserves global information flow. Result: up to 75% less KV-cache memory; up to 6.3x faster decoding at million-token context; beats pure full-attention baselines on short context, long context, and RL scaling.

2.2 Attention Residuals (AttnRes) — Cross-Depth Selective Retrieval

Standard residual connections accumulate uniformly across depth — early-layer representations get diluted in deeper layers. AttnRes adds selective retrieval so the model can pull high-value representations from earlier layers directly across depth. Moonshot reports roughly 25% training efficiency gain with under 2% extra compute.

2.3 Stable LatentMoE — 896 Experts, 16 Active

Combined, Kimi K3 delivers roughly 2.5x overall scaling efficiency vs Kimi K2 — same compute budget, stronger intelligence.

Core benchmarks below are Moonshot AI self-reported (different harness per model: K3 uses Kimi Code, GPT uses Codex, Claude uses Claude Code). Independent third-party reproduction is still in progress.

Key takeaways:

K3 standard pricing matches Claude Sonnet 5 ($3/$15) but with a context window 5x larger. Cached input drops to $0.30/M (one-tenth of standard); Moonshot reports over 90% cache hit rates in coding workflows, making effective input cost very low. China API: ¥20/M input, ¥100/M output, ¥2/M cached; consumer kimi.com free tier available; prepaid plans from ¥199 (promo through August 11).

Six steps to start using Kimi K3 today:

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..."}]
)

July 27 open-source commitment: Moonshot's official WeChat announcement confirms full model weights release on July 27. K3 will become: the largest downloadable open model to date; the first open weights above 2T parameters; a new fine-tuning baseline for the open community. Training uses MXFP4 weights and MXFP8 activations with quantization-aware design; MXFP4/NVFP4 builds expected on Hugging Face.

Summary: Kimi K3 delivers genuine engineering innovation at the architecture level, matching or beating closed-source flagships on key tracks like long coding tasks and document understanding, with fair pricing and a full open-weight promise — signaling China's open AI ecosystem moving from "price to win" toward "challenge the frontier." Key dates: WAIC releases July 17–20 → July 27 K3 full weight release.

Lay out the alternatives: running Kimi Code / API agents on a personal Mac suffers sleep and network interruptions on long-context jobs; waiting for 7/27 self-hosting requires a 64+ card supernode — not viable for most teams short term; relying on a single closed API misses K3's 1M flat-pricing advantage on context length and cost. For iOS CI/CD, persistent Kimi Code, and 24/7 AI agent production, KVMNODE dedicated Mac Mini M4 cloud rental is usually the better host: Apple Silicon unified memory, open sudo, flexible daily/weekly/monthly terms. See the pricing page, help center, or order directly.

Data as of: July 16, 2026 · Benchmarks are Moonshot AI self-reported · Sources: kimi.com/blog/kimi-k3, API Platform docs, Artificial Analysis, OpenRouter