docs/QWEN_MODEL_POLICY.md

Qwen Model Policy for JackAILocal

JackAILocal uses Qwen3.5 as the practical default family and Qwen3.6-27B as a power profile for high-end machines.

Default pack

The default pack is not the ceiling. It is a sensible offline default designed to run on ordinary laptops without forcing the user to understand model size, quantization, VRAM, context length, or backend selection.

Default candidates:

• Qwen3.5 2B for low-spec machines.
• Qwen3.5 4B for normal laptops.
• Qwen3.5 9B as the default smart offline assistant.
• Qwen3.5 27B when the machine has enough memory.

Power pack

Qwen3.6-27B is included as a power profile candidate.

Use it when:

• the machine has RTX 3090-class hardware or better;
• the machine has roughly 24 GB VRAM or enough RAM for a quantized fallback;
• the user selected the Power package;
• the use case benefits from stronger coding, reasoning, vision, or agentic workflows.

Do not use Qwen3.6-27B as the default for ordinary laptops.

Hackathon rule

The hackathon rule is not "Qwen 9B only". The rule is that the Space must be powered by models under the 32B parameter limit. Qwen3.6-27B is eligible by parameter count. Avoid ambiguous models whose published parameter count is over the limit.

Backend policy

Qwen3.5 models can be used through Ollama when the selected tag is available. They can also be provided as GGUF files for llama.cpp.

Qwen3.6-27B is packaged through its Ollama model reference when that model is selected. A GGUF quantization may also be used for an explicitly supplied local llama.cpp server if the artifact and hash are pinned in the manifest.

Context policy

Do not default to the maximum native context. Long contexts increase memory usage and can cause out-of-memory errors.

USB defaults:

• normal profiles: 4K to 8K context;
• power profiles: 8K to 16K context;
• long-context mode: explicit advanced option only.

Product wording

Use this wording:

JackAILocal ships with a practical Qwen3.5 default pack for ordinary laptops, then unlocks Qwen3.6-27B and other power models when the detected machine can run them.

Do not say:

JackAILocal is limited to Qwen 9B.