Krasis documentation

Choosing a model

Pick a supported checkpoint that suits your machine and your work.

Start with the launcher catalogue

Use the curated download flow in krasis to choose a supported Hugging Face checkpoint. This avoids guessing whether an architecture can run. For a manual download, use the original supported safetensors checkpoint, including its configuration and tokenizer files.

The default model directory is ~/.krasis/models/ on Linux and %USERPROFILE%\.krasis\models on Windows. A model can also live elsewhere and be selected by its path.

Check three kinds of space

• System RAM holds model data and runtime backing storage. Limited VRAM does not remove the need for RAM.
• GPU VRAM holds active weights, attention state and working buffers.
• Disk space must cover both the downloaded source checkpoint and generated Krasis caches.

The hardware checker estimates memory fit. Leave room for your operating system and other applications. Its results are estimates rather than measured speed or quality guarantees.

Match the model to the task

For coding agents, tool-call reliability and usable context matter as well as answer quality. For long documents, check the tested context and leave room for the answer and any tool schemas. For images, check vision support separately from text support.

Read ADQ results for scoped productive-task evidence and benchmarks for measured performance. Qualification belongs to a specific checkpoint and tested profile. Changing attention precision, cache format or other quality-sensitive settings can move outside that evidence.

Keep model and cache files distinct

The source checkpoint is the input; Krasis builds its own optimised caches for running it. A slow first launch is expected. Subsequent compatible runs reuse those caches.

Do not delete caches as a routine response to slow generation: rebuilding costs time and disk activity. If startup reports an incompatible or damaged cache, follow the specific error and consult Troubleshooting.

Next: Memory and speed.