Shrinking a model is about more than file size. Memory use, response quality, latency and maintenance burden change together. The right quantization level depends on the loss your use case can accept.
4-bit is not always the right answer
More aggressive quantization reduces memory use but may lose consistency and nuance in some tasks. Evaluation should use your own data and tasks rather than a generic benchmark.
The real cost drivers
Hardware is only the beginning. Model serving, monitoring, security patches, capacity planning and incident response make up a significant part of operating cost.
- GPU or unified memory capacity
- Concurrent users and token throughput
- Model update and rollback process
- Energy, cooling and operations time
Cloud or on-premise?
Sensitive data, low predictable traffic or offline requirements strengthen the on-premise case. Rapid scaling and experimenting with different models may favor managed APIs.