Longer chains of reasoning
With thinking enabled, the model works through planning, math, debugging and analysis step by step before it answers, and Buble shows a summary of that reasoning next to the reply.
Kimi K3
Kimi K3 Chat for Deep Reasoning, Images and Research
Kimi K3 is Moonshot AI's flagship model, released in July 2026: a sparse mixture-of-experts model with 2.8 trillion parameters, built for long reasoning, coding, visual understanding and agent-style tool use. On Buble you can chat with it in English or Chinese, attach images, let it search the web, and switch thinking on when a problem needs it.
Thinking Mode
Image Input
Web Search
Tool Calling
Coding
English and Chinese
Capabilities
What Kimi K3 adds
Kimi K3 is a large step up in scale over Kimi K2.5. The practical gains show up in tasks that need many steps of reasoning, several sources, or a careful reading of images.
Stronger coding
It writes, explains and refactors code across languages, and it is especially good at front-end work: describe a page or paste a screenshot and ask for the components.
Native vision
Vision is trained into the model rather than added on. Charts, screenshots, handwritten notes, diagrams and photos of documents can be read and reasoned about in the same answer.
Research with live sources
When web search is on, the model looks up current pages, reads them and answers with what it found, which matters for prices, releases, news and anything newer than its training data.
Agent-ready tool use
The model was trained for multi-step tool calls. On Buble it uses that ability to decide when to search; through the API it can call your own functions.
Bilingual by design
Moonshot builds Kimi for Chinese and English. You can ask in one language, paste sources in the other, and request the answer in either.
On Buble
How Kimi K3 works on Buble
These details come from Buble's current chat configuration, so they describe what you get in the chat window today.
Pick Kimi K3 in the model menu
Open Chat, choose the model from the list, and start typing in English, Chinese or a mix of both.
Up to 8 images per message
Attach PNG, JPEG, WebP or GIF files of up to 10 MB each. Images from your latest message are sent with it, so re-attach an earlier image if the next question depends on it.
Up to 200K tokens of context
The model itself reads up to about one million tokens. Buble sends your most recent 30 messages, up to 200,000 tokens, which covers long documents and long working sessions.
Billed by the token
Input costs 400 credits per million tokens and output 2,000 credits per million, with repeated context at 40. A 3,000-token question with a 1,500-token answer costs about 4 credits.
Subscription required
This model is available to subscribers.
Reasoning summaries
When thinking is on, you see a short summary of how the model reached its answer, which makes long replies easier to check.
Use Cases
Where Kimi K3 is worth the extra credits
For short factual questions a smaller model is enough. Reach for Kimi K3 when the cost of a wrong or shallow answer is higher than the cost of the tokens.
Debugging and code review
Paste the failing function, the error and the test, and ask for the cause before the fix. The model is good at following a bug across files when you give it the relevant parts.
Research briefs
Turn on web search, ask a question that needs several sources, and request a short brief with links you can check.
Reading charts and screenshots
Upload a dashboard, a spreadsheet screenshot or a slide and ask what changed, what is missing or what the numbers imply.
Long document analysis
Paste a contract, a paper or meeting notes and ask for risks, open questions or a structured summary, then follow up on any part.
Planning and decisions
Lay out options, constraints and goals, and ask the model to compare them, list trade-offs and say what it would check next.
Translation with context
Translate between Chinese and English while keeping terminology, tone and formatting, and ask for notes on choices that could be read two ways.
Workflow
Getting better answers from Kimi K3
The model responds best to a clear goal and the right material, not to long instructions about how to think.
Step 01
State the outcome
Say what you need at the end: a fix, a table, a decision, a draft of a given length. One sentence is enough.
Step 02
Add the material
Paste the code, text or data, or attach images. Include constraints such as the language version, audience or deadline.
Step 03
Choose thinking and search
Turn thinking on for multi-step problems and search on for anything recent. Leave both off for quick rewrites.
Step 04
Check and iterate
Read the reasoning summary, test the result, and reply with what did not work. Follow-ups in the same thread keep the context.
Model Foundation
What is inside Kimi K3
The figures below come from Moonshot AI's model card on Hugging Face. They explain why the model is strong on long tasks without being slow on every token.
Kimi K3 is a sparse mixture-of-experts model with 2.8 trillion total parameters, of which about 104 billion are active for each token. It has 896 experts and routes every token to 16 of them, so it keeps the knowledge of a very large model while doing a fraction of the work per step.
Its native context window is 1,048,576 tokens. Moonshot trained it with quantization-aware training, using MXFP4 weights and MXFP8 activations, which is why the open weights stay practical to serve despite their size.
The API launched on 16 July 2026, and the weights were published on Hugging Face shortly after under the Kimi K3 License. Text, images and video are handled by the same model; the Buble chat currently accepts text and images.
| Detail | Why it matters |
|---|---|
| 2.8T total / ~104B active parameters | Very broad knowledge with per-token compute closer to a mid-size model. |
| 896 experts, 16 active per token | Fine-grained routing lets different experts specialize in code, math or languages. |
| 1,048,576-token native context | Room for whole codebases or document sets; Buble sends up to 200K per request. |
| Native text, image and video understanding | Visual questions are answered by the same model that does the reasoning. |
| MXFP4 weights, MXFP8 activations | Lower serving cost and faster responses for a model of this size. |
| Open weights under the Kimi K3 License | Teams can self-host, subject to the license terms. |
Benchmark scores published at launch come from Moonshot's own evaluations. Test the model on your own tasks before relying on it for production work.
Comparison
Kimi K3 or Kimi K2.5?
Kimi K2.5 is offline on Buble for now, so treat this table as a reference: the difference is scale and what each model is built for.
| Decision point | Kimi K3 | Kimi K2.5 |
|---|---|---|
| Model size | 2.8T total, ~104B active | 1T total, 32B active |
| Native context | About 1M tokens | 256K tokens |
| Context sent on Buble | Up to 200K tokens | Not available right now |
| Thinking, search, tools, images | Yes | Yes |
| Access | Subscribers | Temporarily offline on Buble |
| Best for | Hard reasoning, code and research | Everyday questions at a much lower cost |
Practical Notes
Kimi K3 limits to know before you rely on it
A larger model still makes mistakes. These notes help you use it where it is strong.
It can be confidently wrong
Check facts, figures and citations, especially for legal, medical or financial questions. Ask for sources when they matter.
Thinking makes answers slower
Reasoning tokens take time and are billed as output. Turn thinking off for simple rewrites, translations and short answers.
Search results vary
Web answers are only as good as the pages found. If a source looks weak, ask the model to search again with a narrower query.
No video upload in chat
The model understands video, but the Buble chat accepts images only. Send key frames as screenshots instead.
Context is trimmed in long threads
After 30 messages or 200K tokens, the oldest turns drop out. Start a new chat with a short summary when a thread gets very long.
Do not paste secrets
Leave passwords, keys and private personal data out of prompts, as you would with any online AI service.
FAQ
Kimi K3 FAQ
Short answers about the model, access and cost on Buble.
Start
Try Kimi K3 on your hardest question
Bring a bug, a research question or a dense document, and see how far a frontier model can take it.
Also on Buble: ChatGPT 5.5 · Gemini 3.1 Pro