keepstar/Kali-Lite-8B

Kali-Lite 8B

Local, frugal and inspectable Anima configuration by Kalicorp.

Kali-Lite 8B is not a foundation model trained from scratch by Kalicorp. It combines the upstream Qwen/Qwen3-8B model with a Kalicorp system doctrine, Ollama configuration and an execution harness designed for local operation.

Provenance

The upstream Qwen3-8B repository is Apache-2.0 licensed. Kalicorp-authored repository material follows the licensing terms of the Kali-Lite project. See the licensing notice before redistributing weights.

Design principles

Kali-Lite is configured to:

• —distinguish facts, hypotheses, proposed actions and observed results;
• —never claim that a command/tool was executed without evidence from the harness;
• —avoid inventing system state;
• —keep the operator in control of privileged or destructive actions;
• —perform inference locally when deployed with a local Ollama runtime;
• —add no Kalicorp telemetry to local inference.

Ollama Modelfile

The exact release configuration is provided in Modelfile.

ollama pull qwen3:8b
ollama create kali-lite -f Modelfile
ollama run kali-lite

What is and is not included

This release package documents the Kalicorp configuration. A Hugging Face repository may be published without redistributing upstream weights. If weights/GGUF files are later added, their exact provenance, quantization and checksum must be documented.

Limitations

Kali-Lite inherits the limitations of its base model and runtime. Small local models can produce plausible but incorrect information, lose constraints in long contexts and make tool-use mistakes. A configured context length is not a guarantee that all information in that context will be used correctly.

Local inference is not automatic compliance with GDPR, the EU AI Act, ANSSI guidance or any certification framework.

Project

Source and security documentation: Kalicorp/kalicorp-kali-lite on GitHub.

Kalicorp — Le Sanctuaire numérique européen