Updated 2026-10-27
What Is Kimi K3?
Kimi K3 is a 2.8-trillion-parameter open-weight model with 1M-token context, vision input, reasoning modes, and API pricing from ¥20/MTok input.
TL;DR
- Kimi K3 is Moonshot AI's previous flagship model: a 2.8-trillion-parameter open-weight model with a 1,048,576-token context window.
- It natively supports vision input, tool calls, structured output, streaming, and low/high/max reasoning effort; thinking is always on.
- Kimi API list price is ¥20 per 1M input tokens and ¥100 per 1M output tokens, before context-cache savings.
- Kimi K3.1 Pro is the newer flagship: a 3-trillion-class model, also with 1M context, priced at ¥13/MTok input and ¥68/MTok output.
What Kimi K3 actually does
Kimi K3 is a large language model designed for long-horizon coding, knowledge work, reasoning, and agentic workflows. According to Moonshot AI, it has 2.8 trillion parameters, uses Kimi Delta Attention and Attention Residuals, and activates 16 of 896 experts through a sparse Mixture-of-Experts design.
For developers, the important practical properties are its 1,048,576-token context window, native vision input, function/tool calling, JSON Schema structured output, partial completion, and automatic context caching. It is available through the Kimi Open Platform API with model ID kimi-k3.
The newer K3.1 Pro adds stronger coding, 3D creation, finance, and long-agent performance. It remains focused on the same developer surface: Chat Completions-style calls, reasoning_effort control, image/video input, tool calls, and a 1M-token context.
Kimi K3 capabilities at a glance
| Feature | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|
| Parameters | 2.8 trillion | 3-trillion class |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Input modalities | Text and vision | Text and vision |
| Reasoning effort | low / high / max; always-on thinking | low / high / max; always-on thinking |
| Input price / MTok | ¥20 | ¥13 |
| Output price / MTok | ¥100 | ¥68 |
| Cache-hit input / MTok | ¥2 | ¥1.30 |
| Weights | Released | Scheduled before Nov 4, 2026 |
Kimi K3 vs. newer and closed models
Moonshot positions K3 as an open-weight model aimed at long engineering sessions and knowledge work. Its published successor, K3.1 Pro, reports 65.7% on Terminal-Bench 4.0, slightly below Claude Opus 5.5 at 66.4% and above GPT-6 Astra at 58.2%.
On Kimi Code Bench 3.0 Long-horizon, Moonshot reports K3.1 Pro at 61.2%, compared with 55.95% for K3. This suggests a meaningful generational gain on multi-step coding-agent tasks, though vendor benchmarks should not be treated as independent verification.
Who should use Kimi K3
- Choose K3.1 Pro for a current flagship API with the lowest published Kimi input/output token price and stronger long-horizon coding claims.
- Choose K3 when you need the already-released K3 weights or want a stable previous-generation model for compatibility testing.
- Look elsewhere if your workload is mostly short, latency-sensitive calls where a smaller model is cheaper, or if you need web-search tooling in production before it stabilizes.
Frequently asked questions
How much does Kimi K3 cost?
On 2026-10-27, Moonshot listed Kimi K3 at ¥20 per 1M input tokens and ¥100 per 1M output tokens. Cache-hit input costs ¥2/MTok; cache writes cost ¥20/MTok for 5-minute TTL or ¥40/MTok for 1-hour TTL.
Is Kimi K3 better than Claude or GPT?
It depends on the task. Moonshot's K3.1 Pro report places it close to Claude Opus 5.5 and ahead of GPT-6 Astra on Terminal-Bench 4.0, but those are vendor-published results. Run the same benchmark on your own prompts and tools before making an architecture decision.
Where can I try Kimi K3?
Official access is available through the Kimi Open Platform at platform.kimi.com. You can also test K3.1 Pro in the K3.1 Studio workbench after signing in at kimik3pro.com.