kimi-k2
Google LLC (Vertex AI)/🇺🇸 US/chat
Kimi K2 Thinking is an open-source model that operates as a "thinking agent," reasoning step-by-step while using tools to achieve state-of-the-art performance on various benchmarks. It is capable of executing up to 200-300 sequential tool calls without human intervention, allowing it to solve complex problems across a wide range of tasks. The model uses Quantization-Aware Training (QAT) to support INT4 inference, which provides a roughly 2x improvement in generation speed.MoreLess
Which id to call
vertex/kimi-k2This exact deployment on Google LLC (Vertex AI), with no routing and no failover. Send it as the model field.
Input /1M
$0.60
$0.06 cached
Output /1M
$2.50
4.2x input
Context
262K
262K output
Added
Jul 2025
chat
Capabilities 4/8
Provider rates
What kimi-k2 costs
Provider prices per 1M tokens, updated September 8, 2026.moreless
These are the upstream provider rates. Pay as you go adds 5%, or 0% if you bring your own keys, and there is no per-request fee. Prompt caching and routing change what you pay against these rates, not the rates themselves.
Input /1M
$0.60
Output /1M
$2.50
Cache write /1M
$2.50
Cache read /1M
$0.06
What a workload costs
- 100K input + 10K output
- $0.0850
- 1M input + 100K output
- $0.85
- 10M input + 1M output
- $8.50
At the rates above, before caching. A cache read costs $0.06 per 1M, so repeated context lands under these figures.
OpenAI compatible
Call it in three lines
Change the base url, use your Requesty key, set the model to vertex/kimi-k2. 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="vertex/kimi-k2", messages=[ {"role": "user", "content": "Explain quantum computing in one paragraph."}, ], ) print(response.choices[0].message.content)
Reference
Specs and data terms
What the catalog reports for this deployment, and what the provider does with the traffic.
Released 2025-11-06
Benchmark scores
Benchmarks measure the model, not this endpoint, so they are the same wherever these weights are served.
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 official model cards, Artificial Analysis and public leaderboards. They measure specific skills and do not capture every aspect of model quality, so test on your own workload.
Same provider
More from Google LLC (Vertex AI)
Newest first, on the same provider and the same key.
Reference
kimi-k2 questions
How much does kimi-k2 cost?
kimi-k2 is priced at $0.60 per million input tokens and $2.50 per million output tokens when accessed via Requesty. Prompt caching is supported, which can cut effective input cost by up to 90% on repeated context. Those are the upstream provider rates: pay as you go adds 5% on top, or 0% if you bring your own provider keys.
What is the context window of kimi-k2?
kimi-k2 has a context window of 262K tokens, with a maximum output of 262K tokens per response. That's roughly 350 words of input you can fit in a single prompt.
How does kimi-k2 perform on benchmarks?
kimi-k2 scores 94.7% on Math Index, 94.7% on AIME 2025, 93.0% on τ²-Bench. See the full benchmark chart above for results across MMLU Pro, GPQA Diamond, SWE-Bench Verified, HumanEval, MATH, AIME, MMMU, and LiveBench.
What can kimi-k2 do?
kimi-k2 supports tool calling, extended reasoning, prompt caching, structured outputs (JSON schema). You can call it through any OpenAI-compatible client by pointing base_url to Requesty.
How do I use kimi-k2 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 "vertex/kimi-k2". The Quickstart above shows Python, JavaScript and cURL snippets.
Can I run kimi-k2 through Requesty?
Yes. kimi-k2 runs through Requesty's OpenAI-compatible API, served from Google LLC (Vertex AI). You do not host the model yourself: point base_url at Requesty, set the model to "vertex/kimi-k2", and requests are routed to the upstream provider with automatic failover. The same key gives you 600+ other models too.
Call kimi-k2 through one endpoint
One key for this endpoint and 600+ other models. No markup on provider prices, automatic failover, caching built in.