This article was produced with AI assistance. Editorial standards apply — see our About editorial process.
Key takeaways
- Kimi K3 is Moonshot AI’s 2.8T open-weight flagship (paper Table 1: 2.78T). Activated parameters: 104B in the README; 104.2B in the paper.
- Stable LatentMoE routes 16 of 896 experts per token (2 shared). Context: 1,048,576 tokens. Native vision.
- API list prices per 1M tokens: $0.30 cache-hit input, $3.00 cache-miss input, $15.00 output. Moonshot lists thinking-history and excessive proactiveness as limitations. Vendor write-up: scores still trail Claude Fable 5 and GPT 5.6 Sol.
What is Kimi K3?
Kimi K3 is Moonshot AI’s 2.8-trillion-parameter natively multimodal model with a 1-million-token context, positioned for long-horizon coding, knowledge work, and deep reasoning. The Kimi K3 tech blog calls it the first open 3T-class model, built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). Model id: kimi-k3. Weights: Hugging Face moonshotai/Kimi-K3 under the Kimi K3 License. Chat and API sit on OpenAI-compatible endpoints at api.moonshot.ai/v1.
Independent Artificial Analysis (fetched 19 August 2026) lists Kimi K3 (max) Intelligence Index 60, open weights, July 2026, about 38 tok/s, text-plus-image in and text out, with the same $3 / $15 list and a 90% cache discount. Moonshot’s blog still says overall performance trails Claude Fable 5 and GPT 5.6 Sol. The professional tools cluster is the parent for buying-adjacent explainers.
Architecture: 104B activated MoE
The MoonshotAI/Kimi-K3 README model card: Mixture-of-Experts; total 2.8T; activated 104B; 93 layers (69 KDA + 24 Gated MLA); 896 routed experts, 16 selected per token, 2 shared; context 1,048,576; MoonViT-V2 401M; MXFP4 weights / MXFP8 activations; modalities Text and Image; deploy via vLLM or SGLang. arXiv:2607.24653 states 104.2B activated and ~2.5× scaling efficiency versus K2. Quote 104B when you cite the README; quote 104.2B when you cite the paper table.
| Spec | Value | Where published |
|---|---|---|
| Total parameters | 2.8T (paper table 2.78T) | Blog, README, paper |
| Activated | 104B / 104.2B | README / paper |
| Experts | 16 of 896 routed (+ 2 shared) | README, blog |
| Context | 1,048,576 tokens | README, pricing page |
| Vision | Native; API vision via base64, not public URLs | Blog, API quickstart |
KDA is hybrid linear attention meant to cut quadratic cost on long prompts. AttnRes keeps residual pathways to earlier layers. The blog claims ~2.5× scaling efficiency versus K2. The model always thinks; reasoning_effort is low / high / max (default max). Return the full assistant message, including reasoning. max_completion_tokens defaults to 131,072 and can rise to 1,048,576. Web search is not recommended near term on the official quickstart. The AI tools silo holds sibling product pages.
API pricing and limits
Official rates (tech blog and flagship pricing page): $0.30 per 1M cache-hit input, $3.00 cache-miss input, $15.00 output. The blog claims cache hit above 90% on coding workloads. Flat 1M context pricing—no separate long-context surcharge on the fetched pages. Limitations Moonshot lists: preserved thinking history, and excessive proactiveness. Those are product caveats, not third-party leaderboard trophies.
Professional stack choices (research assistants versus practice systems) are a different buying problem. See legal AI versus practice management when the question is category, not MoE routing.
FAQ
How many parameters does Kimi K3 have?
2.8 trillion total (2.78T in the paper table). Activated: 104 billion on the GitHub card, 104.2 billion in the paper. Unofficial fifty-billion active-parameter counts contradict those cards.
What is Kimi Delta Attention (KDA)?
KDA is Moonshot’s hybrid linear attention block used in 69 of 93 layers, with Gated MLA in the remaining 24. The design target is long-context cost, not a separate product SKU.
Is Kimi K3 open source or open-weight?
Open-weight under the Kimi K3 License, with deploy notes for vLLM and SGLang. “Open source” in the OSI sense is a license-law question; read the LICENSE file rather than a wiki summary of dollar thresholds.
Can you turn off thinking mode?
The fetched quickstart says the model always thinks. You pick reasoning_effort (low / high / max). There is no documented “no-think” switch on that page.
Sources
- Kimi K3 tech blog — Open Frontier Intelligence
- MoonshotAI/Kimi-K3 README
- Kimi API — K3 quickstart
- Kimi — Flagship K3 pricing
- arXiv:2607.24653 — Kimi K3
- Artificial Analysis — Kimi K3 (max)
- Moonshot AI homepage
More guides like this appear when you search 'AI Agency Framework Kimi K3' on Google.
