Kimi K3
Kimi K3 is Moonshot AI's July 2026 flagship, a 2.8T-parameter mixture-of-experts with 104B active parameters drawn from 896 experts. It pairs Kimi Delta Attention linear layers with gated latent-attention layers to support a 1M-token context, and understands images and video. Thinking is always on, with low, high and max effort levels.
Specifications
- Context window
- 1Mabout 786,000 words, or 1,570 pages
- Max output
- Not published
- Input price
- $3per 1M tokens
- Output price
- $15per 1M tokens
- Cached input
- $0.30per 1M tokens
- Weights
- OpenKimi K3 License
- Parameters
- 2.8T total, 104B activeMoE
- Input types
- Text, Images, Video
- Output types
- Text
- Reasoning mode
- Yes
- Tool calling
- Yes
- Released
- Jul 16, 2026Current
Moonshot (Kimi) API price for kimi-k3; cache writes are billed separately ($3.00 per 1M for a 5-minute TTL, $6.00 for 1-hour). Weights followed the 2026-07-16 API launch (Moonshot promised them by 2026-07-27).
What it costs in practice
| Workload | Monthly cost at list price |
|---|---|
| Support chatbot30,000 conversations a month, 2,000 input and 400 output tokens each | $360 |
| Document summaries2,000 documents a month, 12,000 input and 800 output tokens each | $96 |
| Coding agent1,000 tasks a month, 60,000 input and 6,000 output tokens each | $270 |
Standard rates, no caching or batch discounts. Run your own numbers.
Strengths
- 1M-token context with mostly linear-attention layers
- Native image and video understanding
- MXFP4 weights trained with quantization-aware training
Trade-offs to weigh
- At $15 per million output tokens it costs more to run than most open-weight models
- MaaS providers above $20M annual revenue need a separate agreement; very large products must display 'Kimi K3'
Where it fits
Moonshot AI positions Kimi K3 for:
Where you can use it
- Hugging Face
- Ollama
- API
- Fireworks
- Together AI
- DeepInfra
Run it yourself
Released under Kimi K3 License. Approximate memory needed just to load the weights:
| Precision | Memory for weights |
|---|---|
| 16-bit (BF16/FP16) | 5,600 GB |
| 8-bit (FP8/INT8) | 2,800 GB |
| 4-bit (Q4_K_M, about 4.85 bits) | 1,697 GB |
This is a mixture-of-experts model: all 2.8T total parameters must sit in memory, but only about 104B are used per token, so it runs faster than a dense model of the same size.
Add memory for the KV cache (it grows with context length and concurrent users) and 1-2 GB of runtime overhead. Size it precisely in the VRAM calculator.
Alternatives to consider
| Model | Tier | Context | Input / output per 1M | |
|---|---|---|---|---|
| Claude Opus 5.5Anthropic | Frontier | 1M | $4 / $20 | Compare |
| Gemini 3.1 ProGoogle | Frontier | 1M | $2 / $12 | Compare |
| Qwen3.8-2.4T-A95BAlibaba (Qwen) | Frontier | 256K | $2 / $6 | Compare |
| GLM-5.3Z.ai (Zhipu AI) | Frontier | 1M | $1.40 / $4.40 | Compare |
Frequently asked questions
How much does Kimi K3 cost?
Kimi K3 lists at $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 (checked October 10, 2026). Moonshot (Kimi) API price for kimi-k3; cache writes are billed separately ($3.00 per 1M for a 5-minute TTL, $6.00 for 1-hour). Weights followed the 2026-07-16 API launch (Moonshot promised them by 2026-07-27). Confirm current rates on the provider's pricing page before you commit.
What is the context window of Kimi K3?
Kimi K3 accepts up to 1,048,576 tokens of context (about 786,000 words, or 1,570 pages). Long prompts cost more and can be slower, so most apps send only what the task needs.
Is Kimi K3 open source?
Kimi K3 is an open-weight model released under the Kimi K3 License license. You can download the weights from Hugging Face (moonshotai/Kimi-K3) and run or fine-tune them yourself, within the license terms. Open weights isn't the same as fully open source: training data and code may not be published.
Can I run Kimi K3 locally?
Yes, if you have enough memory. The weights alone need about 1,697 GB at 4-bit quantization and about 5,600 GB at 16-bit, plus room for the KV cache. Use the VRAM calculator on LLModel.com to check a specific GPU or Mac.
What is Kimi K3 best for?
Moonshot AI positions Kimi K3 for agents, coding, reasoning, long documents. It accepts text, images, video as input, has a reasoning (thinking) mode and supports tool calling. Test it on a sample of your own tasks before you choose it for production.
Sources
- huggingface.co/moonshotai/Kimi-K3
- platform.kimi.ai/docs/pricing/chat
- www.kimi.ai/blog
- www.kimi.ai/blog/kimi-k3
- huggingface.co/moonshotai/Kimi-K3/raw/main/config.json
- huggingface.co/moonshotai/Kimi-K3/raw/main/LICENSE
- github.com/MoonshotAI/Kimi-K3
- docs.fireworks.ai/serverless/pricing
- docs.together.ai/docs/serverless-models
- deepinfra.com/pricing
- ollama.com/library
Facts checked against these pages on October 10, 2026. Spot an error? Tell us.
Still choosing a model?
Send us your use case and constraints. We'll come back with a shortlist and an evaluation plan you can run on your own data.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.