On this page
Kimi K3.1 Pro is open-weight, so you can use it without Moonshot ever seeing your prompts, but at 2.8 trillion parameters its smallest file is 593 GB and needs a server with eight NVIDIA H100 GPUs. For most people, private Kimi K3.1 means an API host that doesn’t keep prompts: DeepInfra and Fireworks don’t by default, and Ollama Cloud says zero retention. Kimi’s own app and API keep your input while your account is open and use it to train, and for anything truly sensitive, Off Grid AI runs a smaller model on your own device for free.
Download Off Grid AI Desktop (free, Mac, Windows and Linux)
What Kimi K3.1 is
The open model is called Kimi K3.1 Pro. Moonshot AI’s model card lists 2.8T total parameters with 104B active per token, text and image input, and a context of 1,048,576 tokens. It’s post-trained from Kimi K3 and leads Moonshot’s own tables on coding, for example 65.7 on Terminal-Bench 4.0 against 42.3 for DeepSeek V4.1 Flash.
The Kimi K3.1 Pro License lets you use, change and sell products built on it. Two limits apply to big companies. A company selling it as an API with more than $20 million in revenue over 12 months needs a separate agreement with Moonshot. A product with over 100 million monthly users or $20 million in monthly revenue has to show “Kimi K3.1-Pro” in its interface.
Why the Kimi app isn’t private
Moonshot’s platform privacy policy says your input is kept while your account is active, on servers in Singapore, and that user content is used to train and refine its models.
That rules it out for client files, code under NDA or anything personal. Send those prompts to a host that doesn’t keep them, or keep them on your own machine.
Private ways to use Kimi K3.1
Option 1: A host that doesn’t keep prompts
Several hosts run the open weights. Prices vary, so compare them, and pick on privacy too.
| Host | Price per 1M tokens (in / out) | Stores prompts? | Trains on them? |
|---|---|---|---|
| DeepInfra | $1.90 / $9.50 | No | No |
| Fireworks | $2 / $10 | No, zero retention by default | Not stated |
| Ollama Cloud | Ollama’s cloud plans | Never logged | No |
| Together | $2 / $10 | Yes, by default | Off by default |
| OpenRouter | From $0.24 / $8.40 | Zero-retention routing is a setting | No |
| Moonshot (Kimi API) | $2 / $10 | Yes, while your account is active | Yes |
Host prices are as listed on OpenRouter, checked 16 November 2026. The cheapest OpenRouter providers serve 4-bit (FP4) builds. Ollama Cloud runs it as kimi-k3.1-pro:cloud, and its Pro plan is $20 a month with $60 of usage credit.
DeepInfra is the one we’d pick for privacy today. It’s also the cheapest of the hosts with a clear no-store policy.
Option 2: Your own server
This is the only fully private route, and it’s a data-centre job. The full model is about 1.56 TB. The smallest community build, KarLexAI’s KLX-IQ1_S, is 592.7 GB and runs on stock llama.cpp. KarLexAI says it fits entirely in one 8×H100 80 GB node with 131K tokens of context, at about 19 tokens a second.
| Build | Size | Where it fits |
|---|---|---|
| KLX-IQ1_S | 592.7 GB | One 8×H100 80 GB node, all on the GPUs |
| KLX-IQ1_M | 647.9 GB | 8×H100 with some experts in system RAM |
| KLX-IQ2_XXS | 709.9 GB | 8×H100 with more experts in system RAM |
| Official weights | about 1.56 TB | Multi-GPU servers running vLLM, SGLang or TokenSpeed |
Sizes are from KarLexAI’s file list, checked 16 November 2026. Upstream llama.cpp added Kimi K3.1 Pro text support on 6 November and vision on 10 November. Moonshot recommends vLLM, SGLang or TokenSpeed for the full weights.
If your company runs one, you can chat with it from Off Grid AI Desktop. Open Settings, find Remote model server, click Add server, type its Address, click Test connection, pick the model under Text + Vision and Save. Any server with an OpenAI-compatible /v1 API works, and the model shows in the chat with a Remote tag. The steps are in how to use a remote llama.cpp server from Off Grid AI Desktop.
Option 3: Off Grid AI Boost (check the list at launch)
Off Grid AI Boost is a flat-price API for open models on our own GPUs. The plan list names Qwen 3.8 and DeepSeek V4.1 Flash “and more”. Kimi K3.1 isn’t named, so check the Boost model list at launch. DeepSeek V4.1 Flash is on the list today, though it scores lower than Kimi K3.1 Pro on Moonshot’s coding tables.
Boost starts at $4.99 a month for the first cohort, with regular prices from $9.99 to $199.99. We don’t store your prompts or train on them. That’s our promise today, and the full policy comes with launch. Usage limits for each plan aren’t published yet.
Boost starts in January; pre-order to lock in half price. API keys arrive by January 2027.
For truly sensitive work, keep it on your device
Any API, private or not, means your text leaves your computer. If that’s the line, run a smaller open model locally. It won’t match Kimi K3.1 Pro on hard coding tasks, but it works offline and nothing leaves the machine.
| Your device | What to run in Off Grid AI | File |
|---|---|---|
| 16 GB laptop | Qwen 3.8 27B as a 2-bit build | 7.21 GB |
| 32 GB computer | Qwen 3.8 27B, UD-Q4_K_M | 16.5 GB |
| Phone | Qwen 3.5 in Off Grid AI Mobile | Depends on the phone |
On a computer:
- Install Off Grid AI Desktop and open Models, then Text.
- Type Qwen3.8-27B in Search models…. Results from Hugging Face show under More from Hugging Face.
- Pick a file. Each one shows its size and whether it’s Tight or Too big for this computer.
- Choose Download, then Use when it finishes.
On a computer, the app marks a file up to 65% of your RAM as Runs well and up to 82% as Tight. For phones, see how to run Qwen 3.8 on your iPhone or on Android.
What goes wrong, and the fix
| Symptom | Check | Fix |
|---|---|---|
| Multi-turn chats lose track over the API | Whether you send back the full assistant message | Moonshot says to pass back reasoning_content and tool_calls, not only content |
| Answers take a long time | Reasoning effort | Set reasoning_effort to "low" or "high"; the default is "max" |
| A cheap host gives worse answers | Its quantization on OpenRouter | Pick a provider serving a higher-precision build |
| You’re not sure a host keeps nothing | Its own privacy page | Use DeepInfra, Fireworks or Ollama Cloud, or turn on zero-retention routing in OpenRouter |
FAQ
Is Kimi K3.1 open source?
The weights are open. Moonshot AI publishes Kimi K3.1 Pro on Hugging Face under its own Kimi K3.1 Pro License, which allows free use and changes. Companies selling it as an API with over $20 million in yearly revenue need a separate agreement.
Can I run Kimi K3.1 on my own computer?
Not on a normal one. The model has 2.8 trillion parameters and the smallest community GGUF is 592.7 GB, built to fit one server with eight 80 GB NVIDIA H100 GPUs. Even a 512 GB Mac Studio can’t hold it.
Does the Kimi API keep my prompts?
Moonshot’s platform privacy policy says it keeps your input while your account is active, stores data on servers in Singapore, and uses user content to train and refine its models.
Which Kimi K3.1 API is most private?
DeepInfra says it doesn’t store or train on your data, Fireworks says zero data retention by default, and Ollama Cloud says prompts are never logged or trained on. Together stores prompts by default.
Does Off Grid AI Boost serve Kimi K3.1?
Boost names Qwen 3.8 and DeepSeek V4.1 Flash on its plan list, not Kimi. Check the Boost model list at launch. Boost is a pre-order with API keys by January 2027.
Related guides
- Best Private LLM APIs in 2026 (Prompts Never Stored or Trained On)
- How to Run DeepSeek V4.1 Flash in 2026 (What It Really Takes, and the Private Option)
- How to Use DeepSeek Privately on Your Mac or Windows PC in 2026
- How to Use a Remote Ollama or llama.cpp Server From a Desktop AI App on Your Mac in 2026
Keep your most sensitive work on your own machine: download Off Grid AI Desktop and try Qwen 3.8 27B today.