kimi-k3
Moonshot AI/Open weights/7 providers
Kimi’s most capable flagship model to date, with 2.8 trillion parameters. It is built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals, with native visual understanding and a 1M-token context window. It is the world’s first open-source model in the 3-trillion-parameter class, designed for frontier intelligence scenarios including long-horizon coding, knowledge work, and reasoning.MoreLess
Which id to call
kimi-k3Requesty routes this id across every provider serving kimi-k3, picking on price and health and failing over automatically. The id stays valid when a provider changes underneath it.
From /1M input
$3.00
Fireworks AI
Context
1.0M
1.0M output
Best first token
286ms
sference
Regions
2
217 tok/s best
Capabilities 5/8
-↑ 20/s
Week of Sep 28, latest, median of 7 providerslog axis
Median tokens per second once generation starts, measured on traffic routed through Requesty through Oct 2. Hover a week to read it.moreless
The axis is logarithmic, because the lines here are more than 12x apart and a linear axis would press most of them flat. Each gridline is a step in ratio rather than a fixed amount. Each provider line is that provider's own median across every model it serves, which is the grain published per provider, so a provider line can cover weeks this model was not called in yet. The exact per-provider figures for this model are ranked below.
7 endpoints / 2 regions
Providers serving kimi-k3
Provider prices, per 1M tokens. Pay as you go adds 5%, or 0% on your own keys.moreless
The two panes rank kimi-k3 as measured on Requesty traffic: by the provider that served it, and by the region the request came from, where the share is how much of that region's traffic this model carries. Whole-window figures, because that is the grain Requesty publishes per provider and per region.
By provider
Median tokens per second once generation starts. Bar length is the value, so the longest bar wins.moreless
Whole-window figures as published, not an average of weekly points. The axis stops at 112/s: one row runs past it and is notched there, because stretching the axis to the outlier flattens every other bar. The figure beside each bar is exact either way.
By traffic region
Median tokens per second once generation starts. Bar length is the value, so the longest bar wins.moreless
Whole-window figures as published, not an average of weekly points.
| Rank | Cost 10wk | Privacy | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 |  Fireworks AI | Global | 1M | $3.00 | $15.00 | $0.30 | 2.01s | 51.0 | 99.8% | zdr | |
| 2 |  Lyceum Technology | EU | 1.0M | $3.00 | $15.00 | $0.75 | 2.06s | 217.0 | 97.0% | zdr | |
| 3 | Moonshot AI | Global | 1.0M | $3.00 | $15.00 | $0.30 | 5.66s | 37.0 | - | zdr | |
| 4 |  Nebius AI | EU | 1.0M | $3.00 | $15.00 | - | 1.24s | 59.0 | 99.4% | zdr | |
| 5 |  sference | EU | 1.0M | $3.00 | $15.00 | $0.45 | 286ms | - | 99.6% | zdr | |
| 6 | Tencent | Global | 1.0M | $3.00 | $15.00 | $0.30 | 9.04s | 14.0 | - | - | no-train |
| 7 |  TensorX Ltd. | EU | 1.0M | $3.00 | $15.00 | $0.75 | 1.34s | 40.0 | 99.3% | zdr |
A column is blank where no qualifying sample exists, and every row links to that provider's endpoint page.
model=
Which id to call
One base url, https://router.requesty.ai/v1, and one key for every id here.
kimi-k3Requesty routes this id across every provider serving kimi-k3, picking on price and health and failing over automatically. The id stays valid when a provider changes underneath it.
OpenAI compatible
Call it in three lines
Change the base url, use your Requesty key, set the model to any id on the left. Existing OpenAI SDK code needs no other edit.
123456789101112131415from openai import OpenAI client = OpenAI( api_key="YOUR_REQUESTY_API_KEY", base_url="https://router.requesty.ai/v1", ) response = client.chat.completions.create( model="kimi-k3", messages=[ {"role": "user", "content": "Explain quantum computing in one paragraph."}, ], ) print(response.choices[0].message.content)
Artificial Analysis
Benchmark scores
Artificial Analysis Coding Index: a composite of coding evaluations including LiveCodeBench, SciCode and Terminal-Bench.
Graduate-level physics, chemistry & biology questions designed to resist Googling.
Artificial Analysis Intelligence Index: a composite of multiple evaluations measuring overall model capability.
Scores from artificialanalysis.ai. They measure the model, not the provider, so they are the same on every endpoint above.
6
More from Moonshot AI
Reference
kimi-k3 questions
Which providers serve kimi-k3?
kimi-k3 is available from 7 providers through Requesty: Fireworks AI, Lyceum Technology, Moonshot AI, Nebius AI, sference, Tencent. All endpoints share one OpenAI-compatible API and one key, and Requesty fails over between them automatically.
How much does kimi-k3 cost?
Pricing starts at $3.00 per million input tokens and $15.00 per million output tokens on Fireworks AI, the cheapest endpoint. Prices vary by provider and region; the table above shows every endpoint. Those are the upstream provider rates: pay as you go adds 5% on top, or 0% if you bring your own provider keys.
Is kimi-k3 open weights?
Yes. kimi-k3 is an open-weights model from Moonshot AI, which is why multiple inference providers can host it. Provider choice affects price, latency, and data-privacy terms, all compared above.
What is the context window of kimi-k3?
kimi-k3 supports up to 1.0M tokens of context, with up to 1.0M output tokens per response. Some providers expose a smaller window; check the per-endpoint context column above.
Which provider is fastest for kimi-k3?
In recent production traffic through Requesty, sference had the lowest median time to first token (286ms) for kimi-k3. Latency shifts over time, so Requesty's latency-based routing picks the fastest healthy provider per request automatically.
How do I use kimi-k3 with the OpenAI SDK?
Install the OpenAI SDK, set base_url to "https://router.requesty.ai/v1", set your API key to your Requesty key, and set the model to "kimi-k3" or any endpoint id from the table above. The quickstart section shows Python, JavaScript and cURL snippets.
Should I call kimi-k3 by its managed id or a provider id?
Use the managed id, "kimi-k3". Requesty picks which of the 7 providers serves each request based on price and health, and fails over when one degrades, so the id keeps working while the routing changes underneath it. There is also an EU-only id, "kimi-k3@eu", which routes the same way but only through EU-hosted providers. Send a provider id like "fireworks/kimi-k3" only when you need one specific deployment and want no failover.
Route kimi-k3 through one endpoint
One key for 7 providers on this model and 600+ others. No markup on provider prices, automatic failover, caching built in.