Skip to content
C-Kernel-Engine docs Antshiv Open Source

C-KERNEL-ENGINE · GUIDE

Start with Gemma or Qwen3

Gemma 3 270M IT and Qwen3 0.6B both answered a question on a Linux laptop. Pick either for a first run. The 270M in Gemma's name means 270 million parameters. The Q5_K_M Gemma file was about 248 MiB here. The Q8_0 Qwen file was about 610 MiB. Allow extra disk space for the generated runtime.

Choose one model

From the CKE checkout after setting up Linux, run one of these pairs in your terminal. Keep that terminal open for the next command.

Examplebash
MODEL=hf://unsloth/gemma-3-270m-it-GGUF/gemma-3-270m-it-Q5_K_M.gguf
RUN="$HOME/cke-runs/gemma3-270m-1k"
Examplebash
MODEL=hf://Qwen/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf
RUN="$HOME/cke-runs/qwen3-06b-1k"

Gemma model card · Qwen3 GGUF and model card

Start a conversation

Examplebash
version/v8/scripts/cks-v8-run run "$MODEL" --run "$RUN" \
  --context-len 1024 --max-tokens 128 --temperature 0

Without a prompt argument the runner opens an interactive terminal (REPL). Ask “What is the capital of France? Answer in one word.” Type /exit when finished. For scripts, add --prompt 'What is the capital of France? Answer in one word.' to ask once and exit.

Give an application the same model

Examplebash
.venv/bin/python -m pip install -r server/requirements.txt
version/v8/scripts/cks-v8-run serve "$MODEL" --run "$RUN" \
  --no-build --host 127.0.0.1 --port 8080 --model-name cke-local

Leave the server running. In a second terminal, check health, discover the model ID and request your first answer. If you want to start directly with HTTP, omit --no-build on the first launch and specify --context-len 1024. That lets the server build the bundle.

What worked here

On CKE revision 95d9f28af we built each server bundle in a fresh directory from its cached GGUF. Both returned “Paris” in one output token. The rendered question used 21 input tokens with Gemma and 24 with Qwen. In the terminal REPL, Gemma answered and stopped on its end token. Qwen3 used the 128-token allowance while reasoning; its terminal output was longer than its HTTP reply. For a coding agent, pick a model with a recorded tool task in the compatibility matrix.

An old Qwen cache lacked its chat template. An old Gemma bundle loaded and passed health, then failed with an undefined beginning-of-sequence token during the first request. Building in fresh directories fixed both cases and preserved the older builds. If you see the same symptoms, rebuild from the matching checkpoint before changing sampling settings.