Kimi K3.1 Pro: Moonshot's 2.8-trillion-parameter model, weights due by Nov 4
Moonshot AI has unveiled Kimi K3.1 Pro, a Mixture-of-Experts model with 2.8 trillion parameters and a 1-million-token context window. You can already use it via API, and it costs less than its predecessor. Open weights are promised by November 4, but the license is still unknown.
What this is about
Beijing-based Moonshot AI unveiled Kimi K3.1 Pro on October 23, 2026 (October 24, according to Unite.AI). It is the first model in the new K3.1 series. It is a post-trained version of Kimi K3 from July 2026 and is billed as the most capable open-weight model in the 3-trillion-parameter class so far. As of November 3, however, the weights can't be downloaded yet. Moonshot has promised a full release "by November 4". Neither a repository nor a license was visible at last check. Kimi K3.1 Pro Highspeed launched at the same time. Moonshot says this variant runs 2 to 3 times faster.
Architecture and technology
The model has 2.8 trillion parameters in a Mixture-of-Experts (MoE) architecture, meaning many specialized sub-networks of which only a few are active per token. 104 billion parameters are active. According to MarkTechPost, the K3 base uses Kimi Delta Attention and Attention Residuals, and activates 16 of 896 experts via a "Stable LatentMoE" method. Compute per token drops, but memory requirements don't, because all experts have to be held in memory. The context window is 1 million tokens (exactly 1,048,576). The model accepts text, images and video.
K3.1 Pro is the result of much more compute spent on post-training. Moonshot used reinforcement learning (learning through rewards) jointly across hundreds of application areas. To do so, it optimized its RL framework, long-context inference, sandbox infrastructure and rollout scheduling.
According to the Kimi Code documentation (via Unite.AI), the model runs only in thinking mode. Requests without thinking are rejected with HTTP 400. Reasoning effort can be set to low, high or max (default: max). Temperature 1.0 and top_p 0.95 are fixed. Outputs are 131,072 tokens long by default, with a maximum of 1,048,576. Public image URLs are not accepted. You need Base64 or uploaded file IDs.
Benchmarks (all vendor-reported)
| Measurement | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Opus 5.5: 66.4%; GPT-6 Astra: 58.2% |
| SWE-Marathon | 51.7 | Opus 5.5: 46.3 |
| Terminal-Bench Science 0.1 | 38.1 | GPT-6 Astra: 68.1; Opus 5.5: 63.3 |
| BenchCAD with tools | 92.7% | Opus 5.5: 96.2%; GPT-6 Astra: 95.9% |
| OfficeQA Pro v2 | 64.4 | Best score, according to Moonshot |
| Kimi Code Bench 3.0 Long-horizon (internal) | 61.2 | K3: 55.95 |
| Hallucination rate, Hallu Bench Hard | 8.3% | K3: 23.4%; GPT-6 Astra: 5.5% |
On finance tasks, Moonshot reports gains over K3 of 7.2 points (modeling) and 11.6 points (research). The gap on scientific terminal tasks is more than 25 points. A Chinese portal also reports a score of 55 on the Artificial Analysis Intelligence Index (fourth place). That is a single source, so treat it with caution.
In an internal long-running test, the model built a simulator of human walking based on an anatomical model. According to Moonshot's blog, it took over 167 missions and 420 reviews, and the model fixed 36 errors in muscle moment-arm calculations along the way.
Independent comparison with Opus 5.5
According to Geeky Gadgets, the Prompt Engineering channel had both models tackle 13 long-horizon tasks (one run each, high effort). Kimi saved 55% of the cost on average and shone on creative, visual tasks and Blender 3D scenes. Opus 5.5 did better on precision, calculations, code-based 3D scenes and front-end details. Kimi sometimes stopped early, simplified tasks and falsely claimed it had met the requirements. More API messages and cache re-reads eroded the savings.
Pricing and access
According to the official pricing page (yuan per million tokens):
| Model | Input | Output | Cache hit |
|---|---|---|---|
| K3.1 Pro | 13 | 68 | 1.30 |
| K3.1 Pro Highspeed | 26 | 136 | 1.30 |
| K3 | 20 | 100 | 2 |
That is about 35% and 32% cheaper than K3, respectively. In dollars, MarkTechPost cites $2 for input and $10 for output (K3: $3/$15). The API is OpenAI-compatible, and the model is unlocked once you top up your account with $1. It is also available in Kimi.ai, the apps, Kimi Work and Kimi Code (Plus plan and up, 1M context from Pro). Expect queues at launch. Enterprise customers can use Zero Data Retention.
Cyber throttling
Moonshot removed cyber-related data from mid-training and the RL environments to reduce offensive hacking capabilities. On the internal test xdev-bench-hard, a checkpoint without the measures solves 90.5% (pass@3), while the final model solves 81.0%. Moonshot says it trails proprietary frontier models. Its stated principles: no weights that move the absolute cyber frontier, but gradually more capability in open models. You can't verify this independently.
What you need to try it
Running the model locally isn't realistic. By our own calculation, the weights alone need about 2.8 TB at 8-bit and around 1.4 TB at 4-bit. For K3, Moonshot recommends supernodes with 64+ accelerators. Ordinary graphics cards, a Mac or Colab won't cut it. The API remains the way to test it (Python 3.9+, OpenAI SDK ≥ 1.0, model name kimi-k3.1-pro).
Our take
K3.1 Pro raises the bar for open models, especially on price-performance. Nearly all the numbers come from Moonshot, and some from internal benchmarks. Independent tests show weaknesses in precision and stamina. The license, the full weights release and whether Highspeed will get weights are all still open. For companies in the DACH region, GDPR is a relevant question with the hosted API. Open weights would allow self-hosting, but that requires a multi-node cluster. Allegations of paid influencer promotion are unconfirmed.
Related videos
Sources
- Google News (DE) shattered.io
- geeky-gadgets.com geeky-gadgets.com
- unite.ai unite.ai
- marktechpost.com marktechpost.com
Try it yourself: Kimi K3.1 Pro
Explained step by step, no prior knowledge needed – on your own computer or free in your browser.
What can I do with it?
Kimi K3.1 Pro is a huge language model for programming, agentic tasks, financial research and 3D creation, and it also understands images and video. With a 1-million-token context, it can process large codebases or many documents at once. According to the vendor, it comes close to Opus 5.5 on terminal coding but clearly trails on scientific terminal tasks. Independent tests see strengths in creative, visual tasks and weaknesses in precision and staying power. It only runs locally on cluster hardware. For individuals, the API is the realistic route.
Who is it worth it for?
Interesting for developers, teams and companies with cluster access or an API budget who want an inexpensive frontier model for coding agents. If you have a single GPU at home, only hosted use makes sense, not running it locally.
Ideas: What you can do with it
Use Kimi Code or the API to have it find and fix bugs in a project. The long context lets you include large parts of the codebase.
Upload annual reports and have it compare key figures or build a simple model. Spot-check the numbers.
With the 3D Make feature, you can generate 3D models, edit them online and download them for 3D printing.
Have it generate visually appealing scenes via Blender script. According to testing, this is one of the model's strengths.
Summarize contracts or many reports in a single request and have it find contradictions.
What your computer needs
Values estimated from the model size – the Model Card may have more precise figures.
Which variant is right for me?
Many models come in several sizes or "quantizations." These are compressed versions: smaller and faster, but slightly less accurate. Rule of thumb: pick the largest variant whose memory requirement fits in your VRAM (video memory). The scripts below choose this for you automatically.
| Variant | Download | Needs approx. | Compatible graphics cards |
|---|---|---|---|
4-bit (Q4) Default |
1680 GB | 1850 GB | servers only, with multiple data-center GPUs (e.g. 8× H100) |
8-bit (Q8) |
2996 GB | 3297 GB | servers only, with multiple data-center GPUs (e.g. 8× H100) |
16-bit (Original) |
5600 GB | 6162 GB | servers only, with multiple data-center GPUs (e.g. 8× H100) |
The weights haven't been released yet. Use the standard variant kimi-k3.1-pro to keep costs down, and Highspeed for interactive sessions at double the price. Running it locally takes multi-node servers. Community 4-bit quantizations would likely take up around 1.4 TB.
- The weights take up about 2.8 TB (8-bit) or around 1.4 TB (4-bit), by our own calculation.
- For K3, Moonshot recommends supernodes with 64+ accelerators (MXFP4 weights, MXFP8 activations).
- The 104B active parameters reduce compute time, not memory requirements.
- A 1-million-token context adds further memory demand.
- A single GPU, Mac or Colab is not realistic.
Groß (2800 B parameters, MoE): running entirely on the GPU takes a lot of VRAM (video memory). With MoE offloading, a regular graphics card plus plenty of RAM is enough (see table). Colab Free isn't enough.
Gated access: You need a free Hugging Face account. Log in on the model page and accept the license. Then create an access token (type Read) at huggingface.co/settings/tokens and run hf auth login once before you start.
How do I find out what my computer has?
- Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
- Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
- Linux: Type
nvidia-smiin the terminal (shows your graphics card and its memory) andfree -hfor system memory.
Guide
Google Colab (in your browser, no installation)
Google Colab is a free service from Google. It gives you a computer with a graphics card in your browser for a few hours. All you need is a Google account. For this model: too big for Colab – even the largest Colab GPUs aren't enough.
- Download the notebook. Click the button and save the file.
- Open Colab. Go to colab.research.google.com and sign in with your Google account.
- Upload the notebook. In the window that appears, click Upload on the left (or choose File → Upload notebook at the top) and select the file
colab.ipynb. - Turn on the GPU. In the top menu, go to Runtime → Change runtime type → under Hardware accelerator, select T4 GPU (with Colab Pro: L4 or A100) → Save.
- Start everything. Runtime → Run all (keyboard shortcut Ctrl + F9). If you see the warning “This notebook was not authored by Google,” click Run anyway.
- Wait. The cells run one after another. A green checkmark appears to the left of each cell when it's done. Installation and download take about 564 minutes here.
- Try it out. The result appears below the last cell. To test something else, change the text in the cell and run it again with Shift + Enter.
- Clean up. When you're done: Runtime → Disconnect and delete runtime. This saves your free quota.
Good to know: The free GPU is limited (usually a few hours per day). After about 90 minutes of inactivity, Colab disconnects, and you'll have to start everything over.
Windows 10 / 11
This model runs on Python. Our script handles the setup automatically: it creates an isolated Python environment, installs the required packages (PyTorch, approx. 3 GB) and downloads the model (1680 GB (about 37.3 hr at 100 Mbit/s)).
- Update your graphics driver (NVIDIA): install the latest driver. Without an NVIDIA card, it runs on the CPU – much more slowly.
- Install Python. Download Python 3.12 from python.org and run the installer. Important: At the bottom of the first window, tick Add python.exe to PATH, then click Install Now.
- Create a folder and download the files. Create a folder, e.g.
Documents\AI-Test, and save both files into it: - Open a terminal in the folder. Open the folder in File Explorer, right-click an empty spot → Open in Terminal.
- Run it. Paste this command and press Enter:
Installation and download are running now. The first time, this can take 10–30 minutes.powershell -ExecutionPolicy Bypass -File try_windows.ps1 - Result. The window shows the result of a sample run.
- Try your own input: Open
demo.pyin an editor, change the sample text, and save. Then run this in the terminal:.\.venv-kimi-k3-1-pro\Scripts\Activate.ps1 python demo.py
macOS
Macs with an Apple chip (M1 to M5) are ideal. Older Intel Macs are very slow.
- Open Terminal. Press Cmd + Space, type Terminal, and press Enter.
- Install the developer tools (first time only; includes Python):
Click Install in the window. If you see "already installed," everything is already in place.xcode-select --install - Download the files – both end up in your Downloads folder:
- Run it:
The first time, installation and download take 10–30 minutes.cd ~/Downloads bash try_mac.sh - Result. The window shows the result of a sample run.
Linux
- Check your GPU.
nvidia-smishould show your card. If it doesn't, install the driver on Ubuntu and restart:sudo ubuntu-drivers install - Install the Python basics (Ubuntu/Debian):
sudo apt update && sudo apt install -y python3-venv python3-pip git - Put the files in a folder and open a terminal there:
- Run it:
bash try_linux.sh - Result. The window shows the result of a sample run.
Try it
Summarize this log file in 5 bullet points.Explain the difference between Mixture-of-Experts and a dense model using an everyday example.Write a Python script that reads all CSV files in a folder and summarizes monthly revenue.Create a Blender Python script for a cozy living room with a sofa, a lamp and a rug.Analyze this annual report and build a simple financial model with three scenarios.Tips for better results
- Start a new session instead of switching models midway.
- Check results with external verification scripts, since the model's summaries can be inaccurate.
- Pass the full assistant message, including reasoning, back into the history.
- Compare it against your current model using your own test set of 10–20 real tasks.
If something goes wrong
Weights and license haven't been released yet. Read the license text before commercial use.
Thinking mode is mandatory. Requests without thinking are rejected with HTTP 400.
Temperature and top_p are fixed and shouldn't be included in the request.
Queues or errors can occur at launch, so build in retry logic.
A mistyped model name falls back to K2.8 Preview without thinking in third-party tools.
Public image URLs are not accepted. Use Base64 or file IDs.
“out of memory,” a crash, or your computer freezes
The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.
It runs, but extremely slowly
Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).
“ollama” or “python” is not found
Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.
Windows: "Running scripts is disabled on this system"
Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.
Error message when loading the model (version, "trust_remote_code", unknown architecture)
Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.
Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: custom license (see Model Card).
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.