Kaliko Windows · end of 2026

Kaliko Workspace

The Spark for the lower end.

Local AI on the PC you already have. A 4 GB graphics card is enough. Your models, your data, and your rules for how they run.

See what it does How it fits on 4 GB
Kaliko Workspace chat. A 4-billion-parameter Qwen model running locally through Ollama rewrites an email to be shorter and kinder, and lists what it changed.

Kaliko Workspace answering with a 4B model through Ollama, on the machine it runs on.

Runs your models through Ollama/LM Studio/llama.cpp/vLLM Cloud providers are optional.
4 GBof video memory is enough to chat, build and automate.
0accounts. Install it and start, with no sign-up.
15/15web questions right from a 4B model, two rounds running.
10/12test apps built correctly by a 7B coding model.

Figures are from Kaliko's own tests on a 4 GB card.

System intelligence

It knows your machine before you pick a model.

Kaliko reads your GPU, your RAM and every model's real size, including the memory its context window needs. Then it tells you plainly how each model will run. Nothing is loaded to find out.

  • Runs wellFits in the GPU's memory. Full speed.
  • Runs, slowerSplit between GPU and RAM. It works, just slower.
  • Too bigWould page to disk and freeze the PC, so you're told before you click.
Kaliko's model picker. It reads 'This machine: GeForce GTX 1650, 4 GB VRAM, 16 GB RAM' and marks deepseek-r1 1.5B and Llama 3.2 3B as Runs well, and Qwen 3.5 4B and Qwen 2.5 Coder 7B as Slower.

The Spark for the lower end

RTX Spark is a new PC. Kaliko is for the one you have.

NVIDIA built RTX Spark so a personal AI computer could sit on a desk. Kaliko brings the same idea to hardware people already own: local models that are tuned, checked and kept on track until a small GPU is genuinely useful.

NVIDIA RTX Spark

A new superchip

A Grace CPU and Blackwell GPU in one package, in new laptops and desktops.

Memory
Up to 128 GB, unified
You need
A new machine

Kaliko Workspace

Software that makes small hardware count

Model fit checks, right-sized context windows, and answers checked against their sources so small models stay accurate.

Memory
From a 4 GB graphics card
You need
The PC you have
A starting set for a 4 GB card
ModelDownloadUsed forOn 4 GB
Qwen 3.5 4B reasoning2.7 GBEveryday chat and web questionsAbout 90% on the GPU
qwen2.5-coder:7b4.7 GBBuilding appsPartly in RAM, slower
deepseek-r1:1.5b1.1 GBThe fastest optionFully on the GPU

RTX Spark is a trademark of NVIDIA Corporation. Kaliko is not affiliated with or endorsed by NVIDIA.

Privacy and control

Your models. Your data. Your rules.

01

Private by default

Models, chats and files stay on your computer. Nothing leaves unless you connect a cloud provider yourself.

  • No account, no sign-up
  • Keeps working offline
  • Cloud keys stay encrypted on disk
02

You decide how it runs

Choose the model for every chat and the runtime underneath it, and set how much context it gets.

  • Pick a model per chat, or let several answer
  • Set the context window to fit your card
  • Mix local and cloud however you want
03

Nothing happens behind your back

Anything that sends, deletes or changes something stops and asks first. Every change can be undone.

  • Approvals that survive a restart
  • One-click undo for files an agent wrote
  • Every turn traced, step by step

Every turn, traced

See what the model actually did.

Each request is planned, run and verified by Kaliko's runtime. Open any turn to read its full trace, from the goal to the final check.

Kaliko's runtime trace for one chat turn: goal submitted, plan created, task scheduled and running, agent finished using 520 tokens, task verified and succeeded.

What it does

One workspace. Turn the dial.

Drag, scroll, or use the arrow keys

Kaliko01 / 06

Chat

Your PC is enough.

Kaliko Workspace comes to Windows through the Microsoft Store at the end of 2026.

Back to top