Summary
The Kimi K3.1 Pro launch happened on October 23, 2026. Moonshot AI launched Kimi K3.1 Pro as a large open-weight AI model. It has a 1-million-token context window. It is live now in the Kimi app and API. Moonshot plans to release its full open weights by November 4.
What Happened
Moonshot AI is a Chinese AI company. It makes the Kimi chatbot and coding tools.
On October 23, 2026, Moonshot launched Kimi K3.1 Pro. It calls it the first model in its new K3.1 series. It also calls it its most capable 3T-class open-weight model to date.
The model is live now. You can use it on Kimi.ai, in the Kimi mobile apps, in Kimi Work, and in Kimi Code. Developers can also use it through the Kimi API.
A faster version launched at the same time. It is called Kimi K3.1 Pro Highspeed. Moonshot says it runs 2 to 3 times faster than the standard model.
The full open weights are not out yet. Moonshot plans to release them by November 4, 2026. That date is a plan, not a done deal. We will know more when the files actually appear.
Key Details of the Release
Here are the facts, plain and simple:
Very large model, smart design. Kimi K3.1 Pro has 2.8 trillion parameters in total. Only 104 billion are active for each token. This design is called mixture-of-experts. It sends each task to a small group of expert parts. That is how a huge model can still run fast.
1-million-token context. The model can read up to 1 million tokens in one prompt. That is a very long document, a large codebase, or many files at once. It takes text and images as input.
Lower API price. The API costs $2 per million input tokens. Output costs $10 per million. Cached input costs $0.20 per million. That is 33% less than the older Kimi K3. K3 cost $3 for input and $15 for output. The Highspeed version costs twice as much: $4 for input and $20 for output.
More Key Details
Full-project coding. Moonshot says the model can handle full coding projects. It can fix small bugs in real code. In addition, it can build full projects in game tools like Unreal, Unity, and Godot. It can make 3D models with Blender and CAD tools.
Better finance work. Moonshot says the model is more reliable for finance tasks. It can run full finance workflows from start to finish. On one internal test, its claim-level error rate fell. It went from 23.4% on K3 to 8.3% on K3.1 Pro. Moonshot puts it another way: finance errors fell by about 65%. These are Moonshot’s own test results. No outside lab has checked them yet.
Strong scores on one coding test. On Terminal-Bench 4.0, Kimi K3.1 Pro scored 65.7%. That is above GPT-6 Astra at 58.2%. It is just below Claude Opus 5.5 at 66.4%. This test checks coding agents on real terminal tasks. Again, this table comes from Moonshot. Treat it as a company claim, not an independent verdict.
One weak spot stands out. On Moonshot’s own science chart, K3.1 Pro scored 38.1%. That is far below GPT-6 Astra at 68.1%. So it is not the best pick for hard science work.
One independent score looks good too. Artificial Analysis is an independent test site. It gave Kimi K3.1 Pro a score of 55 on its Intelligence Index. That puts it fourth in the world. Only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full rank above it. The older K3 scored 44. So the jump is real on this outside test as well.
The Safety Twist: Hacking Skills Removed on Purpose
This launch has an unusual safety story. Moonshot removed cyber-related data from the model’s training. It did this on purpose, before the open release.
Moonshot tested two versions of the model. One version had no safety filters. It scored 90.5% on an internal cyber test. The public K3.1 Pro scored 81.0%. Moonshot cut that skill down on purpose.
Moonshot said, “We aim not to release open weights that would extend the absolute frontier of cyber capability.” In plain words: once weights are public, anyone can download them. Moonshot does not want its open model to make hacking much easier.
This matters because open models work differently from closed ones. With a closed model, the company controls the servers. With open weights, users run the model on their own machines. Safety filters in the app do not travel with the files. That is why Moonshot acted at the training stage, not just in the app.
Some experts still raise a question. A skilled user could fine-tune the open weights later. That could bring back some of the filtered skill. Moonshot has not answered that point in full. It is fair to keep that doubt in mind.
What Open Weights Mean for You
Open weights are the trained model files. When Moonshot releases them on November 4, developers can download them. They can run the model on their own hardware. They do not have to depend on Moonshot’s servers.
But be realistic about size. This is a 2.8-trillion-parameter model. It will need very strong hardware to run in full. Most students and freelancers will not run it on a normal laptop. The practical way to try it is the Kimi app or the API.
The API price cut helps here. At $2 per million input tokens, long coding and research tasks cost less than before. For a freelancer who builds client tools or checks large files, that price drop is real money saved.
This is the second big open-weight launch in a week. We also covered the MiniMax M3.1 Flash open-weights launch, which took a very different price path.
Why It Matters for Freelancers and Students
Kimi K3.1 Pro is not just another chatbot update. It points to three useful shifts.
First, top-level AI is getting cheaper. A model that ranks fourth in the world now costs $2 per million input tokens. A year ago, that level of skill cost far more. Small teams can now afford tools that only big firms could use before.
Second, long context changes how you work. With 1 million tokens, you can feed in a lot at once. Add a full client brief, past work, and reference files in one go. The model can keep the full picture in mind. That helps with long reports, large code projects, and deep research.
Third, open models give you choice. If Moonshot releases the weights as planned, developers can build on the model without asking permission each time. They can also check how it behaves on their own systems. That choice matters for privacy and for long-term projects.
For students, the free route is simple. Try the Kimi app first. Test it on coding homework, long reading, or a research plan. Compare its answers with Claude or ChatGPT. Do not trust any single model on facts. Check key claims yourself.
For freelancers, the finance and coding gains matter most. If you build tools or write code, test K3.1 Pro first. Try it on a small paid job. The same goes for finance reports. Company benchmarks are a guide, not proof.
If you want a cheaper open model to compare, see our DeepSeek V4.1 Flash update from earlier this month.
What to Watch Next
Three dates and questions matter now.
One, will the open weights really arrive by November 4? Watch for the actual files and the license terms. The license will decide who can use the model for business work.
Two, will outside labs confirm the benchmark scores? Independent tests in the next few weeks will show if K3.1 Pro truly stands near Claude Opus 5.5.
Three, will the safety choice hold up? Other labs are watching this closely. If Moonshot’s approach works, more open models may follow the same path. If not, the debate over open weights will grow louder.
The Bottom Line on the Kimi K3.1 Pro Launch
Kimi K3.1 Pro is a strong, fresh launch. It is fast, cheaper than its predecessor, and close to the best closed models on several tests. Its safety choice sets it apart from most open releases.
But two honest notes remain. Most benchmark scores come from Moonshot itself. And the open weights are promised, not yet delivered. Try the live model. Judge it on your own tasks. Wait for the November 4 release. Only then call it a true open model.