kimi-k3
Kimi K3 · Moonshot
×2.1
serving
- Multiplier
- ×2.1
- Context served
- 1.00M tokens
- Max output
- 32.8K tokens
in ×1.0 · cached ×0.1 · out ×5
kimi-k3
Kimi K3 from Moonshot, served on an OpenAI-compatible endpoint. Everything below is either the row that will serve your next request or something we measured against the live API.
Published
What you get
The same record /models renders, read from the row that serves the next request for this slug.
Call it
Send a request
Swap the base URL for https://api.tium.ai/v1 and use this slug as the model name. Nothing else about your code changes. Setup for specific tools is on the connect pages.
curl https://api.tium.ai/v1/chat/completions \
-H "Authorization: Bearer $TIUM_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"kimi-k3","messages":[{"role":"user","content":"hello"}]}'Measured
What we found by testing it
Behaviour that is not in the provider's documentation because nobody wrote it down. Each page names the probe and the date.
Do reasoning effort levels actually do anything?
On GLM, yes, measurably. On DeepSeek and Kimi the parameter is accepted and the output does not change.
Where do these models report cached prompt tokens?
All five report cache hits under usage.prompt_tokens_details.cached_tokens, and that is the field we bill from.
Is there a /v1/responses endpoint?
No. A POST to /v1/responses returns 404. Use /v1/chat/completions.
Why did my temperature setting disappear?
Some upstreams allow exactly one value for a sampling parameter. We drop those rather than let the request fail, and name what we dropped in a response header.
Why does my answer come back empty on a reasoning model?
Because max_tokens budgets the reasoning too, and the reasoning is emitted first. Every model here returned an empty answer at 512 and below.
Can you send images to these models?
Three of the five see images. GLM-5.3 refuses the request outright. DeepSeek-V4-Pro accepts it, returns 200, and does not see the image.
Billing
What a request costs
Every model is billed in weighted tokens against one multiplier, so the arithmetic is the same whichever you pick. How a weighted token is computed walks through it, and pricing has the tiers.
Cached input is the lever worth knowing about on this model as much as any other. Cache hits are reported back to you in the response body, in the same field we bill from, and on a long agent context they are the difference between most of a request and a fraction of it. See prompt caching.