Private LLM
Your own LLM. A few clicks away.
Run any open model from the App Gallery as your private chat and API, on a GPU your Kinesis Grid finds and keeps running.
Your Kinesis Grid
Summarize this supplier contract in five bullets.
Here are the key terms: a 24-month term, net-60 payment, and…
1 GPU · 48 GiB VRAM
Answered by a GPU in your Kinesis Grid.Nothing went to a model vendor.
Illustrative
Why private
A model of your own beats a model you rent by the token.
Frontier APIs are great for the hardest problems. For everyday work on your own data, a private model gives you control.
- Your dataModel vendor’s API: Prompts and documents go to the vendor, under its terms for keeping them.Your private LLM: Prompts stay with your app, on a GPU in your Kinesis Grid.
- The modelModel vendor’s API: The vendor changes or retires it on its own schedule.Your private LLM: The model you picked, until you choose to switch.
- Making it yoursModel vendor’s API: Only the settings the vendor offers.Your private LLM: Your system prompt, your documents, your fine-tuned weights.
- LimitsModel vendor’s API: Shared rate limits and token quotas.Your private LLM: A GPU of its own, with no token quota.
- Where it runsModel vendor’s API: Wherever the vendor runs it.Your private LLM: The cloud, region or server you choose.
- BillingModel vendor’s API: Per token: the bill grows with every question.Your private LLM: Per GPU-hour, however much your team asks.
Any model
Pick a model, or name any other.
The App Gallery has the leading open models ready to launch, and a picker for the whole Ollama library.
Open LLMs, each with its own private chat
Open WebUI, a ChatGPT-style chat, on Ollama: set up for you, on GPUs in your Grid.

Choose Custom Model
Any model from the Ollama library
Search by name, pick a size up to 130B parameters, and Kinesis sizes the GPU memory for it. Llama, Mistral, Phi and the rest of the library.

gpt-oss:120b
OpenAI gpt-oss
1 GPU · 40 GiB VRAM

gemma4:31b
Google Gemma 4
1 GPU · 48 GiB VRAM

qwen3.6:27b
Qwen 3.6
1 GPU · 48 GiB VRAM

deepseek-r1:32b
DeepSeek-R1
1 GPU · 64 GiB VRAM
CPU trialgemma4:e4b
Google Gemma 4
1 GPU · 16 GiB VRAM
CPU trialqwen3.5:9b
Qwen 3.5
1 GPU · 16 GiB VRAM
Also in the gallery: Microsoft BitNet on llama.cpp, which needs no GPU. Models marked CPU trial can also start on serverless CPU; larger ones run on a GPU in your Grid.
Switch models when a better one comes out: Gemma 4, Qwen 3.5 and Qwen 3.6 joined the gallery as they were released.
Not in the gallery? Bring the model.
Deploy any container image, or build one from a GitHub repository or a Dockerfile. It runs on the same Grid, with the same HTTPS address, logs and metrics.
- vLLM + Hugging Face
- llama.cpp
- Your fine-tuned weights
Chat and API
A chat for your team. An API for your code.
Every model comes with a ChatGPT-style chat and an OpenAI-compatible API, each on its own HTTPS address. Existing code only changes its base URL.
Chat for your team
https://your-app.apps.kinesiscloud.com
Open WebUI, a ChatGPT-style chat in the browser.
API for your code
https://your-app.apps.kinesiscloud.com:<port>/v1
OpenAI-compatible: point your SDK or tool at the new base URL.
$ curl https://your-app…/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "gemma4:31b",
"messages": [{"role": "user",
"content": "Draft a reply to this ticket"}]}'Illustrative
Why Kinesis
Everything around the model is already there.
A private LLM is more than a GPU and a model. Kinesis brings the rest.
The GPU, found for you
Your Kinesis Grid finds a GPU that fits the model: Kinesis capacity, a cloud account you connect, or a server you own.
Live in minutes
No drivers, no gateway, no certificates to set up. The chat and the API come up on their own HTTPS address.
Kept running
If a machine fails or a spot GPU leaves, your Grid moves the model to another machine that fits.
Shared with your team
Models live in your organization’s projects, with SSO, roles, logs and metrics.
Ready for your agents
The REST API and MCP server can launch a model for you, from Claude, ChatGPT or Cursor.
Private all the way down
Need more? Kinesis Private runs the whole platform isolated for you, or fully inside your own data center.
How it works
From a model’s name to your own LLM in four steps.

gemma4:31b
Open WebUI with Ollama
Where it runs
- Your Grid found a GPU that fits
- Model downloaded, chat ready
- HTTPS address live
Illustrative
Choose one from the App Gallery, or name any Ollama model. Each shows the GPU memory and disk it needs.
Your Kinesis Grid: Kinesis capacity, your cloud accounts or your own servers, under the max price per GPU-hour you set.
Quick Launch starts it as it is. Customize First opens every setting: environment variables, hardware and location.
Your Grid finds the GPU and downloads the model. Then the chat and the API answer on their own HTTPS addresses.
Questions
What people ask about private LLMs.
The App Gallery has gpt-oss, Gemma 4, Qwen 3.5 and 3.6, DeepSeek-R1 and Microsoft BitNet ready to launch, and a picker for any model in the Ollama library, up to 130B parameters. For anything else, such as a Hugging Face model on vLLM or your own fine-tune, deploy a container image.
Your prompts go to your own app, on a GPU in your Kinesis Grid, and never to a model vendor. Your Grid never places work on community hardware unless you choose Kinesis Batches. For a deployment isolated for you, or one fully inside your own data center, see Kinesis Private.
No. Your Kinesis Grid can use Kinesis capacity, the cloud accounts you connect with an API key, or servers you own, in any mix. Small models, up to about 9B parameters, can also try out on serverless CPU.
Yes. Each model template also serves the Ollama API on its own HTTPS port, with OpenAI-compatible endpoints under /v1. Point your SDK, agent framework or coding tool at that base URL.
Anyone with the app’s HTTPS address can reach it, and the chat and the API start without a sign-in. Before you share it, limit the app to your own IP addresses, or turn on the chat’s sign-in and put your own gateway in front of the API.
You pay for the GPU your model runs on, by the GPU-hour, never per token, and a spot GPU never costs more than the max price you set. For compute you bring, Kinesis charges a small Grid fee instead. The details are on our pricing page.
Your private LLM is a few clicks away.
Create an account, pick a model, and your Kinesis Grid finds the GPU and keeps it running.