PN — The environmental framing
The environmental framing: fully local execution with 4-bit quantized models (1-5 GB in memory) avoids cloud dependency and reduces energy, with the footprint of large proprietary inference potentially orders of magnitude higher for comparable text-classification tasks.
Source: PUMA project documentation · Traceability: corpus unit
MEM-065· Confidence: verified-at-primary-source