
Kimi K3
Kimi K3 is Moonshot AI's open-weight multimodal model for agentic work. It's a 2.8T-parameter mixture of experts with 104B active per token and a 1M-token context window.
Ways to run Kimi K3
Dedicated endpoint
Get a quote →Your own GPUs, tuned for your traffic. We profile your workload, tune the deployment and put throughput and latency in your contract.
Serverless
API docs →Pay per token on shared GPUs, with no capacity to plan or reserve.
Inside your network
Security →Run a dedicated endpoint inside your own network, including air-gapped environments.
Pricing
Level with the InferenceX benchmark
Total tokens per second per GPU, 16 live agent sessions, same GB300 GPUs
Built on the NVIDIA Dynamo serving stack with vLLM, tensor parallel 8 over two GB300 trays. Your numbers will differ with your prompts and traffic.
How we measured thisUse it with the agents we set up for you
Point any open-source agent at your endpoint, or pick from our catalog and we set it up as part of your contract.
Kimi K3 FAQ
How much does Kimi K3 cost on Saturn Cloud?
Serverless costs $0.75 per 1M input tokens and $3.75 per 1M output tokens. A dedicated endpoint is priced per GPU-hour for the GPUs it runs on, and we size the GPUs to your traffic.
Does Kimi K3 support tool calling?
Yes. In multi-turn and tool-call conversations, send back the full assistant messages, including tool calls and reasoning, in the message history.
Can I turn off thinking in Kimi K3?
No. Thinking is always on. Set reasoning_effort to low, high or max. The default is max.
What sampling settings does Moonshot AI recommend?
Temperature 1.0, with top_p 0.95 for single-step tasks and 1.0 for agentic tasks.
Do you guarantee Kimi K3's performance?
Yes, on dedicated endpoints. Once your workload is profiled and benchmarked, we commit to the throughput and latency agreed for it, backed by contractual remedies such as service credits where applicable.