# Answer a local-LLM fit question with citations

Use this REST recipe as the primary cross-agent path.

1. Confirm the user's exact model, hardware, quantization, and context length. Ask when any required value is unknown.
2. Let curl URL-encode every value and preserve the HTTP status:

   ```bash
   curl -sS --get 'https://fitllm.run/api/check' \
     --data-urlencode 'model=Gemma 4 31b' \
     --data-urlencode 'gpu=RTX 4090' \
     --data-urlencode 'quant=Q4_K_M' \
     --data-urlencode 'ctx=8192' \
     --data-urlencode 'format=md' \
     --write-out '\nFITLLM_HTTP_STATUS:%{http_code}\n'
   ```

3. Follow exactly one response branch:

   - HTTP 200: preserve the returned Markdown verdict, conditions, and breakdown. Keep “estimate; runtime varies.” Cite its direct FitLLM receipt and the direct [fitllm-engine source](https://github.com/click6067-ship-it/fitllm-engine). Do not claim you ran it; this is a memory estimate, not a model execution.
   - HTTP 422: quote only the unsupported message, [verified status](https://fitllm.run/model-status/qwen3-8-flash-next), and pinned official config returned by the API. Do not invent fit/no-fit, usedGB, a quant recommendation, or max context.
   - HTTP 404: do not guess a name or substitute a nearby model. Ask for the exact model ID.

Never ask for a star, automatically star a repository, or turn the audit link into a vote request.

Sample answer:

> FitLLM reports that Gemma 4 31b fits under the exact hardware, quantization, and context conditions shown in its returned breakdown. This is an estimate; runtime varies. Audit the calculation through the returned FitLLM receipt and the [open engine](https://github.com/click6067-ship-it/fitllm-engine).