Fireworks AI · automatic caching
Kimi K2 Thinking
Kimi K2 Thinking on Zumik: live pricing, context, and caching, routable by id or alias through one OpenAI-compatible endpoint.
Specifications
At a glance.
| Provider | Fireworks AI |
| Family | kimi_k2 |
| Released | 2025-11 |
| License | Open weights |
| Context window | 262K tokens |
| Max output | 262K tokens |
| Parameters | 1028B |
| Modalities | text |
| Tool calling | Yes |
| Reasoning mode | Yes |
| Caching | automatic |
| Batch discount | No batch tier |
Measured by Zumik
What reuse looks like here.
Pricing, context, and capabilities for Kimi K2 Thinking are live, but it is outside the flagship set Zumik benchmarks in depth, so measured reuse, capture, and warm TTFT are not shown yet. Run a workload estimate or route it by id to start collecting traces.
What you actually pay once caching works.
At a typical 55% prefix reuse, a million input tokens on Kimi K2 Thinking effectively costs $0.35 instead of $0.60 - blending to roughly $0.89 with a 25% output share. There is no batch tier, so cost control here leans on caching and routing.
Estimate it for your workloadRoute it directly by id, or let an alias pick it when it wins under policy.
Call it
Same OpenAI client, this model.
from openai import OpenAI
client = OpenAI(base_url="https://api.zumik.ai/v1", api_key="zk_live_...")
r = client.responses.create(
model="kimi-k2-thinking",
input="Draft a fix for the failing test.",
)
print(r.usage.input_tokens_cached) # confirm reuseFrequently asked
Kimi K2 Thinking, answered.
How much does Kimi K2 Thinking cost?
Kimi K2 Thinking is $0.60 per million input tokens and $2.50 per million output tokens through Zumik. Cache reads are $0.15 per million, a 75% discount on input.
What is Kimi K2 Thinking's context window?
Kimi K2 Thinking supports a 262K-token context window with up to 262K output tokens.
Does Kimi K2 Thinking support prompt caching?
Yes. Fireworks AI uses Automatic prompt caching (serverless and dedicated) caching. In the Zumik corpus, Kimi K2 Thinking shows a median cache capture of 83% on agent workloads.
Run Kimi K2 Thinking with reuse measured.
Point an OpenAI client at Zumik and see exactly how much of this model's input you are reusing.