DCL-Qwen3-4B-Instruct-2507-GGUF

DCL-Qwen3-4B-Instruct-2507 — GGUF (Q4_K_M, ~2.5 GB)

The laptop build of DCL-Qwen3-4B-Instruct-2507 — a 4B model fine-tuned to author and repair DCL, Russell East's Declarative Capability Language. Read the main model card for usage and limitations — the short version: system prompt Output ONLY the DCL source. No prose, no explanation, no markdown fences., put the DCL vocabulary reference in your prompt with the brief, and always run the output through the DCL compiler.

Frozen-exam numbers for exactly this quantisation: protocol authoring 7/9 vs the stock base's 2/9, every pass a first-attempt clean compile; repair 3/3. Receipts.

Ollama

ollama run hf.co/RichWoollcott/DCL-Qwen3-4B-Instruct-2507-GGUF

or build from the included Modelfile (bakes in the serving prompt + a low temperature):

huggingface-cli download RichWoollcott/DCL-Qwen3-4B-Instruct-2507-GGUF --local-dir dcl-model
cd dcl-model && ollama create dcl-author -f Modelfile
ollama run dcl-author

llama.cpp

llama-server --model DCL-Qwen3-4B-Instruct-2507-Q4_K_M.gguf --jinja -ngl 999

The stock Qwen3-2507 chat template is embedded in the GGUF and renders correctly under llama.cpp's --jinja.

Credit

DCL is by Russell East — why he created it. Base: Qwen/Qwen3-4B-Instruct-2507 (Apache-2.0, LICENSE retained). Fine-tune by Richard Woollcott.