Ollama 0.32, local models, operations, and privacy

Ollama runs supported open models locally or through configured services. The current release baseline is 0.32.5. Its API provides chat, generate, embed, model management, tool calling, structured output, and OpenAI-compatible routes.

Spring integration

<dependency>
  <groupId>org.springframework.ai</groupId>
  <artifactId>spring-ai-starter-model-ollama</artifactId>
</dependency>

spring.ai.ollama:
  base-url: http://ollama:11434
  init:
    pull-model-strategy: never

Pin model names/digests in production. Pull models during an explicit provisioning stage, not every application startup.

Capacity

Model size, quantization, context length, concurrent requests, GPU/CPU memory, KV cache, batch size, and keep-alive determine latency and throughput. A local endpoint is not free: hardware, power, model storage, warm-up, and operations are costs.

Capabilities belong to models

Ollama may expose a tools or JSON API, but the selected model must perform the behavior well. Test schema adherence, parallel tools, long context, multilingual quality, and prompt injection for each model and quantization.

Security

Do not expose port 11434 publicly. Put authentication, authorization, rate limits, request size, and tenant isolation in a gateway. Local inference reduces external data transfer; it does not automatically provide encryption, auditing, deletion, or safe model provenance.

Feynman check

Ollama is a local model kitchen. The menu says what can be served; your hardware determines how fast; your application still checks who may order.