What you will achieve now
A reproducible report of a local Ollama trial, without customer data and without exposing the service.
- 1 Check the service and network
- 2 Run the synthetic case
- 3 Decide with metrics and risks
Version, model, and reviewed license; local address; size; response; latency; tokens per second; and a reasoned decision.
- Distinguish local Ollama, cloud models, and external integrations.
- Measure a concrete task without turning a single run into a benchmark.
- Detect a network exposure and understand why the local API requires control.
- Include license, hardware, energy, support, and review in the real cost.
1. Define a trial that can fail
We will use a made-up company. The only authorized information says the store is open Monday through Friday. The question will be whether it opens on Saturday. A responsible answer should abstain: Not available.
Do not start with contracts, invoices, or real conversations. First check that you know how to run, measure, stop, and explain the flow with synthetic data.
EMPRESA FICTICIA: Nube Clara HECHO AUTORIZADO: La tienda abre de lunes a viernes de 09:00 a 18:00. PREGUNTA: ¿La tienda abre el sábado? RESPUESTA ESPERADA: No consta.
2. Check version, models, and listening address
Official documentation places the local API at http://localhost:11434/api . In addition, the local API does not require authentication. That convenience is precisely why you must not publish it on a LAN or on the Internet without designing another security layer.
ollama --version ollama list ollama ps # macOS o Linux: la salida segura debe mostrar 127.0.0.1:11434 lsof -nP -iTCP:11434 -sTCP:LISTEN
3. Clone and verify the lab
Aulafy’s MIT lab does not download models, does not use keys, and blocks any destination that is not loopback. Its tests use a simulated local server, so they can run even if Ollama is off.
git clone https://github.com/aulafy/taller.git cd taller/cursos/ia-pymes/laboratorios/ollama-piloto-seguro npm run verificar npm run diagnostico
4. Choose a model without accepting a ranking
There is no permanent “best small model.” Choose one you already have or check its official card. Record the exact name and tag, downloaded size, quantization, license, languages, and limitations. Ollama’s MIT license does not make the model weights MIT.
If the model does not fit with enough headroom for the system, the context, and other applications, try a smaller one. Do not buy hardware before measuring a representative task.
# Inspecciona el modelo local y su licencia
curl http://127.0.0.1:11434/api/show -d '{
"model": "gemma3:4b"
}'
# Sustituye el nombre por uno que aparezca en ollama list
OLLAMA_MODEL=gemma3:4b npm run probar5. Read the evidence, not only the response
The generation endpoint returns times in nanoseconds and token counts. The lab converts those figures and also measures perceived time. Separate initial load from generation: a second query may be faster if the model remains in memory.
EJECUCIÓN DE REFERENCIA — NO ES UNA PROMESA equipo: Apple M4; 24 GB de memoria unificada ollama_servidor: 0.32.1 modelo: gemma3:4b tamaño_aproximado: 3.34 GB respuesta: "No consta." coincide: true tiempo_total: 2.08 s tiempo_carga: 1.64 s tokens_salida: 4 tokens_por_segundo: 44.67
6. Decide with an honest matrix
- Keep investigating: the task works, latency is acceptable, and you can build an authorized test suite.
- Reduce the model or the context: there are waits, memory pressure, or the machine freezes.
- Compare a cloud or hybrid option: you need more quality, concurrency, support, or availability than you can operate locally.
- Discard: review costs more than the manual process, you cannot protect the data, or failure has too much impact.
DECISIÓN DEL PILOTO tarea: modelo_y_licencia: equipo_y_memoria: datos_usados: sintéticos / autorizados / minimizados api_confirmada_en_loopback: sí / no latencia_mediana_de_varias_pruebas: casos_correctos: abstenciones_correctas: correcciones_humanas: coste: equipo + energía + soporte + revisión + mantenimiento decisión: ampliar / ajustar / comparar / descartar responsable: próxima_revisión:
Privacy, cloud, and maintenance
Ollama states that it does not receive prompts or data when the model runs locally. For a strictly local mode, its documentation allows disabling cloud features; doing so removes cloud models and web search. Check the configuration and the logs; do not rely only on the tool’s name.
Update with control: record the version, reread the release notes, and rerun the suite after changing Ollama, the model, quantization, the prompt, or the context. The API aims to maintain compatibility, but it is not strictly versioned.
# Configuración oficial para desactivar funciones cloud
# ~/.ollama/server.json
{
"disable_ollama_cloud": true
}
# Alternativa mediante variable antes de iniciar el servidor
OLLAMA_NO_CLOUD=1Official sources consulted
- Introduction and base URL of the API.
- Local and cloud authentication.
- Generation, timings, and counts.
- FAQ: privacy, cloud, and listening address.
- Official hardware support.
- Ollama repository and license.
If you have saved the evidence from this lesson, continue with “Basic GDPR.” If not, repeat the check before moving on.
Learn with a verifiable reference
Aulafy distinguishes stable concepts from data that change—versions, prices, models, and commands. Check the course card for the review date, the verified scope, and the primary sources.
“Editorial review” means structure, claims, and sources have been reviewed. It does not mean every command has been executed: when a technical test exists, it will be marked as such.