July 16, 2026 launch · KDA architecture · 1M context · full benchmarks · API pricing · July 27 open weights
TL;DR: Moonshot AI just released Kimi K3 — the world's largest open-source AI model at 2.8 trillion parameters. It has a 1M token context window, native vision, beats Claude Fable 5 and GPT-5.6 Sol on several coding benchmarks, and costs $3/$15 per million tokens. Full weights drop July 27. Below: architecture, benchmarks, pricing, how to try it, and whether you should switch.
Kimi K3 is a 2.8-trillion-parameter mixture-of-experts (MoE) model from Beijing-based Moonshot AI. It's the world's first open model in the 3T-class, surpassing the previous record holder DeepSeek V4 Pro (1.6T) by nearly 75%.
| Spec | Detail |
|---|---|
| Total Parameters | 2.8 trillion |
| Architecture | Kimi Delta Attention (KDA) + Attention Residuals + Stable LatentMoE |
| Active Experts | 16 out of 896 (sparse MoE) |
| Context Window | 1,048,576 tokens (1M) |
| Input Modalities | Text, image, video |
| API Model ID | kimi-k3 |
| Pricing | $3 / $15 per 1M tokens (input/output) |
| Open Weights | July 27, 2026 |
The model is live on kimi.com, the Kimi mobile app, Kimi Code, and the Moonshot API. You can try it now for free with a Google account — no credit card needed.
Context truncation on long code tasks — 200K closed models force chunking on large repos
Open-weight scale ceiling — prior max was ~1.6T, limiting frontier coding agents
Theoretical vs practical long context — KV cache costs make advertised windows unusable in production
Closed API dependency — no self-host fallback when pricing or policy shifts
The last 18 months were rough for Moonshot AI. DeepSeek's rise eroded their market position. Kimi K3 is a striking comeback:
This isn't a vanity project — it's a fast-growing business making a serious technical statement.
Full attention scales quadratically with context. At 1M tokens, KV cache memory becomes catastrophic. KDA alternates cheap linear-attention layers with full-attention layers in a 3:1 ratio:
Standard residuals dilute early-layer signals across depth. AttnRes enables selective retrieval across depth — pulling high-value representations from earlier layers. Result: ~25% higher training efficiency at under 2% extra compute.
Just 16 of 896 experts activate per forward pass (1.8% sparsity). Moonshot's fixes: Quantile Balancing, Per-Head Muon, SiTU, and Gated MLA. Together: ~2.5× better scaling efficiency vs Kimi K2.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 |
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 |
| Terminal Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 |
| Automation Bench | 30.8 | 29.1 | 29.7 | 27.2 |
| GPQA-Diamond | 93.5 | 92.6 | 94.1 | 91.0 |
| MMMU-Pro (vision) | 81.6 | 81.2 | 83.0 | 78.9 |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 |
SWE Marathon — sustained long-horizon coding — is the headline: K3 leads at 42.0, a 7-point gap over Claude Fable 5. On Artificial Analysis Intelligence Index v4.1, K3 scores 57.1 (4th), behind Fable 5 (59.9) and GPT-5.6 Sol (58.9) — a 2.8-point gap from #1 to #4.
Caveat: Moonshot self-reported benchmarks. Different harnesses (Kimi Code, Codex, Claude Code). Treat as directionally useful, not definitive.
| Model | Input $/1M | Output $/1M | Cache-Hit Input | Context |
|---|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 | 1M tokens |
| Claude Sonnet 5 (standard) | $3.00 | $15.00 | — | 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 |
K3 matches Claude Sonnet 5 standard pricing but gives 5× the context. Kimi Code workflows hit cache at 90%+ via Mooncake split-inference — effective average input cost can drop to roughly $0.55/M tokens in real usage. vs Opus 4.8: similar or better on several benchmarks at 60% input / 40% output cost.
Just chat — kimi.com, Google sign-up, max reasoning effort, no credit card
API — key at platform.kimi.ai, model kimi-k3
OpenRouter — moonshotai/kimi-k3, official pricing, full 1M context
Wait for weights (July 27) — Hugging Face release; production needs 64+ accelerators
Smoke-test — one API call to verify latency and billing before routing production traffic
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 and identify performance bottlenecks..."}
]
)| Use Case | Best Pick | Why |
|---|---|---|
| Long, sustained coding sessions | Kimi K3 | Leads SWE Marathon; 1M context |
| Complex bug fixes in large repos | Claude Fable 5 | FrontierSWE advantage |
| Terminal/tool-heavy agent workflows | GPT-5.6 Sol | Terminal Bench leader |
| Multimodal document analysis | Kimi K3 | Best OmniDocBench; native vision + 1M |
| Cost-sensitive production | DeepSeek V4 Pro | Output at $3.48/M |
| Open-source self-hosting (post 7/27) | Kimi K3 | Most capable open weights |
When weights drop July 27, Kimi K3 becomes the largest downloadable open-source model ever — the first above 2 trillion parameters. Trained with MXFP4 weights and MXFP8 activations for efficient inference. Day-0 support expected in transformers, vLLM, and SGLang.
Dates to bookmark: Right now → try at kimi.com · July 17–20 → WAIC Shanghai · July 27 → Hugging Face weights
Kimi K3 is the most capable open-source AI model ever released. It doesn't win every benchmark — Claude Fable 5 and GPT-5.6 Sol still lead on specific tasks — but it's competitive across the board, outperforms them on long-horizon coding and document understanding, and ships with a 1M-token context at Sonnet-tier pricing. The July 27 open-weight release is the story to watch for the rest of 2026.
Yes on kimi.com with a free account. API is pay-per-token at $3/$15 per 1M tokens.
Weights release July 27, 2026. Production needs 64+ H100-class accelerators — not a laptop LLM.
K3: 2.8T params, 1M context, stronger benchmarks. DeepSeek: far cheaper output ($3.48 vs $15 per 1M).
Yes for whole-codebase calls, long documents, and long-memory agents. Flat pricing makes full-window use practical.
Moonshot says subsequent updates. Currently only max is available at launch.
K3's API is live today; self-hosting waits for July 27 and a GPU supernode most teams won't buy. If you're on Windows/Linux but need a real macOS GUI to run Kimi Code agents, handle keychain prompts, or benchmark against your current Claude/GPT stack, buying a Mac is expensive and SSH won't click system dialogs. VNCMac remote Mac rental gives you hourly VNC access to an isolated node — route K3 API traffic, run long-context coding sessions, and validate benchmarks from this article without a hardware commitment. See Mac rental plans to start.