Kimi K3 is Moonshot AI’s million-token model for coding, reasoning and long-context agent work. It combines a hosted Kimi API launch with an announced open-weight release, making it relevant to teams that want both immediate API access and a potential future self-hosting path.
At a glance
| Fact | Detail |
| --- | --- |
| Provider | Moonshot AI / Kimi |
| Type | Hosted API with announced open weights |
| Best for | Coding, agents, research, long-context analysis |
| Context window | 1,048,576 tokens |
| API available | Yes |
| Local use | Open weights announced; verify current release availability before deployment |
| Last checked | 02 August 2026 |
Our verdict
Kimi K3 is one of the more interesting Chinese model launches for teams that need huge context and serious agent capability without abandoning the prospect of open-weight deployment. The hosted API is the practical option today; treat any local roll-out as a separate infrastructure decision until Moonshot’s current weight-release and licence terms are confirmed.
| Score | Rating |
| --- | --- |
| Overall | ★★★★☆ |
| Coding | ★★★★☆ |
| Agents | ★★★★☆ |
| Research | ★★★★☆ |
| Budget | ★★★★☆ |
Pricing
Standard Kimi API pricing is per million tokens.
| Meter | Price |
| --- | --- |
| Input | $3.00 / million tokens |
| Cached input | $0.30 / million tokens |
| Output | $15.00 / million tokens |
Best for
- Long-context coding and agent workflows
- Research packs and large document collections
- Teams evaluating future open-weight deployment options
- Cost-conscious frontier-class API experiments
Avoid when
- You need a fully confirmed local deployment today
- You need a mature western enterprise support ecosystem
- A smaller low-cost model is sufficient for routine work
What changed
- 02 August 2026: Profile created for Kimi K3.
- 02 August 2026: Hosted API pricing and million-token context checked against Kimi documentation and provider launch material.
- Future weight-release, licence and API changes will be added here.
Related news
- Moonshot’s Kimi K3 makes the open-model race a scale contest
- OpenAI’s GPT-5.6 push is really a pricing war for enterprise AI
- SpaceXAI ships Grok 4.5, Musk calls it an Opus-Class model at half the price
