Kimi K2 Thinking
moonshot/kimi-k2-thinkingK2 with deliberate reasoning, extended thinking over the same long context, released open.
Model overview
Context window
256K
tokens
Input price
—
quoted on request
Output price
—
quoted on request
Weekly volume
—
tokens / week
Open weights
Yes
self-hostable
Capabilities
4
of 11 tags
Description
Kimi K2 Thinking is a reasoning model from Moonshot AI. Webparam does not stock this model yet and can source it on request — tell us what you need it for and we will come back with availability and a price. Moonshot AI’s stated focus is long-context agentic intelligence (Kimi). Its mean it can also be self-hosted under its licence.
Capabilities
- Reasoning
- Spends extra billed tokens thinking before it answers, trading latency and cost for accuracy.
- Tool calling
- Calls functions you define and returns their arguments as structured data, the basis of agents.
- Streaming
- Emits tokens as they are generated, so answers appear progressively rather than all at once.
- Open weights
- The parameters are published, so the model can be inspected, fine-tuned and self-hosted under its licence.
Strengths & weaknesses
Strengths
- Reasoning over 256K context
- Same price as the base model
- Open weights
Weaknesses
- Slow first token under heavy reasoning
- Thinking tokens inflate output cost
Pricing
| Rate | Price | Unit |
|---|---|---|
| Input | — quoted on request | — |
| Output | — quoted on request | — |
This model is not provisioned yet; the rate is quoted on request.
Context window
256Ktokens
covers prompt and response together, so a long input leaves less room for the answer.
Supported features
| Feature | Support |
|---|---|
| Tool calling | Supported |
| JSON mode | Not supported |
| Streaming | Supported |
| Vision input | Not supported |
| Audio input | Not supported |
| Long context | Not supported |
| Open weights | Supported |
| Reasoning | Supported |
| Fine-tunable | Not supported |
| Batch | Not supported |
| Caching | Not supported |
Example use cases
Long-document reasoning
Planning agents
Complex analysis
Comparison
Compare against anything →| Attribute | Kimi K2 Thinking | DeepSeek R1 | Kimi K2 |
|---|---|---|---|
| Context window | 256K | 164K | 256K |
| Input price | R0.00 | R0.00 | R0.00 |
| Output price | — not charged | — not charged | — not charged |
| Tool calling | Supported | Supported | Supported |
| JSON mode | Not supported | Not supported | Supported |
| Vision input | Not supported | Not supported | Not supported |
| Open weights | Supported | Supported | Supported |
| Prompt caching | Not supported | Not supported | Supported |
Best value in each row is highlighted. Blank cells are facts the supplier does not publish, not zeroes.
Provider information
View Moonshot AI →Moonshot AI · 月之暗面 · Beijing, China · est. 2023
Moonshot's Kimi line made long context its signature, and its K2 generation paired that with serious agentic tool use, released with open weights. The lab moves fast and publishes aggressively, making Kimi one of the most-watched model families in the ecosystem.
Recent releases
Documentation
Frequently asked questions
Input is billed at R0.00 per 1M tokens, and there is no separate output charge. Every figure on this page is an illustrative placeholder.
Yes. Kimi K2 Thinking's parameters are published, so it can be downloaded, inspected, fine-tuned and self-hosted under its licence. Self-hosting trades per-token pricing for infrastructure you run.
Prompt and response share one window: Kimi K2 Thinking accepts up to 256K tokens in total, so a long input leaves less room for the answer. The median across reasoning models is 200K. Recall is strongest well inside the limit. Test at your real prompt depth before relying on the ceiling.
DeepSeek R1 is the closest neighbour in the catalogue, with 164K of context at R0.00 per 1M tokens in. Put both in a comparison before wiring a fallback chain, so the trade-off is a decision rather than a default.