Allocate
Kimi K3
Open weightsKimi K3 is a language model from Moonshot AI with a 1M-token context window. Provider list price is $3 per million input tokens and $15 per million output; on Allocate you pay $3.21 and $16.05. The weights are open, so you can fine-tune it and own the result.
Pricing
Prices checked 2026-07-21.
Price against its peers
Provider list prices per M tokens, Kimi K3 against its nearest language peers by price.
What a real workload costs
Take 1,000,000 requests a month at 1,200 input and 350 output tokens each: 1,200M input and 350M output tokens. At list prices that is 1,200 × $3 + 350 × $15 = $8,850 a month. Billed on Allocate it is $9,470.
Kimi K3 is an open-weights model; the catalog does not list its license, so check the lab’s model card before commercial fine-tuning.
Example usage
Point a route at moonshotai/kimi-k3 and the endpoint never changes; swap the model behind it whenever you want.
curl https://api.allocate.network/v1/chat/completions \ -H "Authorization: Bearer $ALLOCATE_KEY" \ -d '{ "model": "moonshotai/kimi-k3", "messages": [{"role": "user", "content": "Summarise the attached contract."}] }'
Common questions
How much does Kimi K3 cost per million tokens?
Provider list price is $3 per million input tokens and $15 per million output tokens. On Allocate you pay $3.21 in and $16.05 out.
What context window does Kimi K3 have?
1,000,000 tokens (1M). At roughly 0.75 words per token, that is about 750k words of English text per request.
What does cached input cost on Kimi K3?
$0.30 per million tokens at list ($0.32 billed). Repeated prompt prefixes, such as a stable system prompt or tool definitions, bill at this rate instead of the full input price.
Can I fine-tune Kimi K3?
Yes. Kimi K3 is an open-weights model; check the lab’s model card for the exact license terms. Read the license terms before fine-tuning for commercial use. On Allocate the trained weights stay inside your boundary and belong to you.
How do I call Kimi K3 on Allocate?
Send moonshotai/kimi-k3 in the model field of the OpenAI-compatible endpoint at api.allocate.network/v1, or point a route name (like prod/support-agent) at it so you can swap the model later without a deploy.