# INTENTION

Build and publish a Hugging Face Gradio Space that demonstrates
`coder3101/gemma-4-31B-it-heretic` without downloading model weights to the
local machine. Serve a directly derived, pre-quantized checkpoint with useful
chat, vision, streaming, and generation controls.

# INTENTION MATCHED

- Uses the directly derived FP8 checkpoint
  `0ppxnhximxr/gemma-4-31B-it-heretic-FP8-Dynamic`.
- Leaves all model-weight downloads to the Hugging Face Space runtime.
- Targets ZeroGPU xlarge because the 34 GB FP8 model plus cache and runtime
  overhead does not fit safely in the 48 GB ZeroGPU size.
- Provides typed text, UTF-8 text-file attachments, image chat, streaming
  output, a 16,384-token input guard, up to 1,536 generated tokens, explicit
  output-limit notices, examples, API/MCP exposure, local launch scripts,
  tests, and a browser settings page.
- Published at
  `https://huggingface.co/spaces/dev-willbird1936/gemma-4-31b-heretic-fp8`.
- Verified the live `/chat` API with a successful ZeroGPU inference.

# INTENTION NOT MATCHED

- None.