Skip to content
KOLDOS
Sign in
KOLDOS

Run KOLDOS on your own computer

KOLDOS runs on your hardware: the model, the tools and the memory. This page covers the systems it targets, what hardware it needs, how installation works and what to expect on first launch.
Release
No public build yet
Downloadable model launch
Planned window: Dec 30, 2026 to Feb 10, 2027A planned window, not a fixed release date.
Version
Pending. Not published yet
Price
Not final. Not final. KOLDOS Local may be offered for free.

Download

No public build has been published yet

Installers will be listed here for each supported operating system with their version, file size, release date and SHA-256 checksum. For questions about KOLDOS Local, contact the team on Discord.
Check your hardware firstSee the roadmap

Operating systems

Support will be announced per system and version.

SystemStatusVersionsArchitectures
WindowsIn developmentWindows 11, the first platformx86-64 with an NVIDIA GPU
LinuxPlannedPending. Linux is in scope. Process handling on POSIX systems has not been verified yet.Pending. Not published yet
macOSPlannedPending. macOS is supported on a best-effort basis.Pending. Not published yet

GPU acceleration

Which hardware KOLDOS Local can use to speed up inference.

  • NVIDIA GPUsllama.cpp with CUDA. Measured on a GeForce RTX 5050 (8 GB) and an RTX 4050 Laptop GPU (6 GB).In development
  • AMD GPUsRadeon RX cards. Not measured with KOLDOS.Pending. Not published yet
  • Intel GPUsArc discrete cards. Integrated Intel graphics share system RAM, so the hardware checker treats them like the CPU.Pending. Not published yet
  • Apple SiliconM-series chips share one pool of unified memory between the CPU and GPU. Not measured with KOLDOS.Pending. Not published yet
  • CPU onlyRunning the main model from system RAM on the CPU. The optional vision model and OCR already run on the CPU.Pending. Not published yet

Requirements

Minimum, recommended and high-end

Official figures will be published after testing on real hardware. Until then, the fields below stay marked as pending instead of guessed.
ComponentMinimumThe lowest configuration that will be supported.RecommendedThe mid-range machine KOLDOS is designed around.High-endFor the largest models and longest contexts.
Operating systemWindows 11All tiers
CPUPending. Not published yetPending. Not published yetPending. Not published yet
RAMPending. Not published yet24 GBPending. Not published yet
GPUNVIDIA GPU with CUDAAll tiers
VRAMPending. No official minimum yet. KOLDOS runs today on a 6 GB laptop GPU with Q5_K_M and a 24K context.8 GB10 GB or more: the runtime profile moves to Q6_K with a 32K context
StoragePending. Not published yetAbout 5.2 GB for the Q5_K_M model file, plus project memory and tracesPending. Not published yet
InternetNot needed for local chat and coding. Needed to download KOLDOS and the model file, and for web search.All tiers
ModelsKOLDOS 7BKOLDOS 7B, Q5_K_MPending. KOLDOS 7B is the only model today. A 22B model is a future plan.

What the numbers depend on

Until official figures exist, this is what the estimate says the models need at 4-bit with a 4K context. It is arithmetic from model size, not a promise.

How the estimate works
ModelTo run fully on a GPUTo run on the CPU onlyModel file
KOLDOS 7BAvailable5.2 GB VRAM8 GB RAM4.0 GB
KOLDOS 22BFuture14.4 GB VRAM18 GB RAM12.4 GB

Installation

From download to first reply

Draft guide

These steps describe the intended flow. Exact screens and file names will be filled in when the first build is published.
  1. 01

    Check what your computer can run

    The hardware checker estimates which model size and quantization fit in your GPU memory and RAM before you download anything.

    Open the hardware checker
  2. 02

    Download the installer

    Installers will be listed on this page for each supported operating system, with their version, file size and SHA-256 checksum.

    Pending. Not published yetNo public build has been published yet.

  3. 03

    Verify the file

    Compare the checksum of the file you downloaded with the one listed here. If they differ, delete the file and download it again.

    Windows (PowerShell)
    Get-FileHash .\<installer-file> -Algorithm SHA256
    macOS
    shasum -a 256 <installer-file>
    Linux
    sha256sum <installer-file>
  4. 04

    Install and open KOLDOS Local

    Run the installer and start the app. Platform-specific steps will be documented with the first build.

    Pending. Not published yetExact steps depend on the installer format, which has not been published.

  5. 05

    Download the model file

    KOLDOS 7B uses a single GGUF file of about 5.2 GB at Q5_K_M. It is the longest step on a slow connection.

    Compare the models
  6. 06

    Set context length and GPU offload

    Start with the configuration suggested below for your hardware. A longer context lets the model read more text at once but uses more memory.

    Graphics card memory

    System RAM

    For an exact GPU, open these values in the full checker.

    Suggested first settings

    Model
    KOLDOS 7B
    Quantization
    6-bit
    Where it runs
    GPU inference
    GPU offload
    All layers
    Context to start with
    Up to 32K

When a connection is needed

Local inference does not need a model server. Getting KOLDOS onto your machine, and features that talk to other services, need a connection. Whether the app works fully offline has not been specified.

Questions about KOLDOS Local:

Discord username

itzkoldkir

Downloading KOLDOS Local and model files
Yes
Chatting with a downloaded model
Not specified. Whether local chat works fully offline has not been specified.
Web search
Yes, when the feature ships
Using the KOLDOS API
Yes
Updates
Not specified. How updates are delivered has not been announced.