Parameter count is only one input
The Phi-4 report describes a particular 14-billion-parameter model. Parameter counts do not establish equivalent quality across tasks or a device’s running memory requirement. Identify the exact model, quantisation, context settings and runtime. Avoid attaching one fixed size threshold to every product called small or on-device.
A weight-only estimate
For a fictional dense model with 7 billion parameters stored at 4 bits each, weight values alone require 7,000,000,000 × 4 ÷ 8 = 3.5 billion bytes, about 3.26 GiB. This excludes quantisation metadata, caches, activations, runtime buffers and the operating system. It is not a claim that a 4 GB device can load and use the model. Actual file and memory measurements may differ.
| Measure separately | Record |
|---|---|
| Model file | Exact download, licence and file size |
| Loaded memory | Runtime and operating system measurements with settings |
| Context | Input length, cache use and response limits |
| Latency | Time to first output and total completion under defined tasks |
| Quality | Accepted outputs, retries and failure categories |
Map the actual data path
Local execution does not automatically mean no network requests, no logs or no cloud features. Check model downloads, updates, telemetry, connectors, application logs and any fallback service. The Ollama documentation describes configurable operation; your surrounding application may add other paths. Record which ones are enabled instead of inferring privacy from a model file on disk.
A fictional offline summarisation test should use a synthetic document, verify the configured network boundary and inspect accessible application logs. A valid summary with an unexpected upload fails the proposed privacy requirement. This is a suggested test, not a claim that Luminesca ran it or certified regulatory compliance.
Compare accepted work, not parameter counts
Use the model evaluation worksheet to hold tasks and acceptance rules constant. Include energy, equipment, support and review in a cost comparison; owned hardware does not make every extra request economically free. Measure your device and workload before deciding whether a local, remote or mixed design meets the requirement.
Sources and scope
- Microsoft authors, Phi-4 technical report — Documents a particular 14-billion-parameter model and evaluation conditions; not every small model.
- Ollama FAQ — Runtime configuration and operational guidance; not a performance guarantee for a reader’s hardware.
Sources checked October 10, 2026. AI-assisted research and drafting. The examples, calculations and worksheets are Luminesca editorial work; they are not measured trials, first-hand reporting or an independent certification. Universal phone speed, privacy, compliance, cost-saving and model-quality claims have been removed. Report a specific correction with its source.