Krasis documentation

Documentation

Run your first local model, connect your apps and understand the settings that matter.

Krasis runs large mixture-of-experts (MoE) language models on NVIDIA GPUs, including models larger than GPU memory. It keeps model data in system RAM and moves the experts it needs to the GPU. Prompt processing and generation use Rust and CUDA.

New to Krasis?

Start with the Getting started guide. It takes you from installation to a first reply, with a short explanation at each step. You do not need to tune advanced settings first.

Already installed? Go straight to connecting your apps.

Everyday guides

• Choosing a model — what fits, what to download and how to read support claims.
• Connecting your apps — chat apps, coding tools and Krasis Manager.
• Memory and speed — RAM, VRAM, context, quantisation and caching in plain language.
• Troubleshooting — startup, memory, connection and answer-quality problems.

Technical reference

• Configuration — saved profiles, config keys, command-line overrides and runtime settings.
• API and authentication — requests, streaming, tools and API keys that can be changed live.
• Full advanced reference on GitHub — the broader flag catalogue, experimental modes and developer commands.

Check requirements and release notes

Krasis supports native Windows, Linux and WSL2 on x86-64 machines with NVIDIA Ampere or newer GPUs. Use the hardware checker on the install page for a rough memory estimate. The source model and generated caches both need disk space; system RAM also matters.

These guides describe stable 1.0.22. Release history explains changes between versions. Benchmarks records measured performance, and ADQ testing describes the scope of model qualification. A model that fits in memory is not automatically qualified for every task or setting.