Kimi K3
Kimi K3 is Moonshot AI's most capable open model (Kimi K3 License, July 2026): a mixture-of-experts network with 2.8T parameters and 104B active per token. It is built for long-horizon autonomous coding across large repositories and terminal tools, and for agentic knowledge work such as deep research; the card reports 88.3 on Terminal-Bench 2.1. Thinking is always on — you choose the effort, not whether it reasons. It is served by a third-party inference provider through our partner, not on Heabsy hardware: the same key and the same invoice, but no zero data retention, and the request may be processed outside the EEA.
model id: kimi-k3
from openai import OpenAI client = OpenAI( api_key="hb-...", base_url="https://api.heabsy.com/v1",) r = client.chat.completions.create( model="kimi-k3", messages=[{"role": "user", "content": "Hello"}],)print(r.choices[0].message.content)
curl https://api.heabsy.com/v1/chat/completions \ -H "Authorization: Bearer $HEABSY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "kimi-k3", "messages": [{"role": "user", "content": "Hello"}] }'
import OpenAI from "openai"; const client = new OpenAI({ apiKey: process.env.HEABSY_API_KEY, baseURL: "https://api.heabsy.com/v1",}); const r = await client.chat.completions.create({ model: "kimi-k3", messages: [{ role: "user", content: "Hello" }],});console.log(r.choices[0].message.content);
Price list, updated 10 October 2026. No minimum spend, no subscription.
About the model
- Developer
- Moonshot AI
- Released
- July 2026
- Licence
- Kimi K3 License
- Architecture
- mixture of experts · 2.8T parameters, 104B active per token
- Input
- text, images
- Reasoning
- always on, effort selectable
- Languages named by the developer
- none named on the model card
Source: the developer's model card, as of 17 September 2026.
Kimi K3: reasoning is always on
Kimi K3 cannot answer without thinking first: the developer offers three effort levels (low, high, max) with max as the default. Every answer therefore carries reasoning tokens — plan output budgets and latency around that.
In multi-turn conversations the card asks for the full previous assistant message, including its reasoning, to be sent back. With 2.8T parameters it is one of the largest open models in the catalogue; for tasks that do not need it, DeepSeek V4 Pro or the cheaper DeepSeek V4 Flash and GLM 5.3 Flash are the obvious alternatives.
Licence: what it means for your company
Kimi K3 License. A custom MIT-style licence with two conditions: a business offering the model as a service with group revenue above US$20 million over 12 months needs a separate agreement with Moonshot AI, and products above 100 million monthly active users or US$20 million monthly revenue must display “Kimi K3”. Internal use is exempt.
A summary of the licence, not legal advice.
Benchmarks reported by the developer
- Terminal-Bench 2.1
- 88.3
- DeepSWE
- 67.5
- BrowseComp
- 91.2
- GPQA Diamond
- 93.5
Figures from the developer's model card, not our measurement. Source: huggingface.co/moonshotai/Kimi-K3
Technical details
- Context window
- 1M tokens
- Max output
- 65,536 tokens
- Throughput limit
- —
- Tokenizer
- Other
- Streaming
- yes
- Open weights
- —
- max_tokens
- 1 … 65536
This model is served by a third-party inference provider through our partner, not on Heabsy hardware. You keep one key and one invoice from an EU company, but there is no zero data retention: the provider handles the request under its own policy, and processing may take place outside the EEA. For personal or regulated data, use a model on Heabsy hardware.
Questions about this model
›Where is Kimi K3 computed?
At a third-party inference provider that serves this model through our partner, outside our own infrastructure. Where exactly the request is computed is not fixed and may be outside the EEA. The contract and the invoice stay with our EU company.
›Does Kimi K3 store prompts and completions?
There is no zero-data-retention guarantee for this model: the request goes to a third-party provider, which handles it under its own policy and may keep logs. For personal or regulated data, use a model on Heabsy hardware.
›How large is the Kimi K3 context window?
1M tokens, with up to 65,536 tokens of output.
›What does Kimi K3 cost?
$4.00 per million input tokens and $20.00 per million output tokens. No minimum spend and no subscription; the same price list covers the API and the invoice.
›Is there a throughput limit on Kimi K3?
The catalog does not report a throughput limit for this model.
›Which languages does Kimi K3 support?
The developer's model card names no languages. Test the model on your own material before relying on it in a particular language.
›Under which licence is Kimi K3 released?
Kimi K3 License. A custom MIT-style licence with two conditions: a business offering the model as a service with group revenue above US$20 million over 12 months needs a separate agreement with Moonshot AI, and products above 100 million monthly active users or US$20 million monthly revenue must display “Kimi K3”. Internal use is exempt.
›Can reasoning be switched off in Kimi K3?
No. Reasoning is always on; only its effort can be chosen, and every answer carries reasoning tokens.
Compare with similar models
| Model | Developer | Parameters | Context | Licence | Input / 1M |
|---|---|---|---|---|---|
| Kimi K3 | Moonshot AI | 2.8T / 104B | 1M | Kimi K3 License | $4.00 |
| DeepSeek V4.1 Flash | DeepSeek | 552B / 8–16 | 1M | MIT | $0.28 |
| GLM 5.3 Flash | Z.ai | 320B / 18B | 1M | MIT | $0.20 |
| DeepSeek V4 Flash | DeepSeek | 284B / 13B | 1M | MIT | $0.28 |
31 more models behind the same key and the same invoice.
Browse the catalog