Skip to content

// installation

Models & inference

Applies to

Product: Server · Audience: Platform Operator

The inference layer is separate from basebox itself: AISRV talks to the language model via an OpenAI-compatible API – wherever it runs. This section keeps four things apart that are easily confused:

Term Meaning Page
Compatible What can run technically: any runtime implementing the OpenAI-compatible API in the form basebox uses Supported inference backends
Tested What basebox has actually verified – with version, backend, hardware, date Tested models
Recommended What basebox suggests for common scenarios – changes with new models Recommended models
Measured Numbers from reference configurations: throughput, latency, long context Benchmarks

A compatible model is not automatically tested; a tested one not automatically recommended. Each page says what it is.

In this section

  1. Inference architecture – where inference runs, how AISRV talks to it, where the data boundary lies
  2. Supported inference backends – vLLM as validated backend, what every endpoint must fulfil, external providers
  3. Configure models – the AISRV_LLM_* settings and their counterparts in the inference runtime
  4. Tested models
  5. Recommended models
  6. Benchmarks
  7. Using other models – checking, validating and reporting a model outside the lists
  8. LLM recommendations (sizing reference) – which model fits which GPU

The essentials in three sentences

The GPU question is an inference question. Model, quantisation, context length and concurrency determine VRAM – not basebox. The platform itself needs no GPU; the service models (RAG, OCR, STT) need their own predictable capacity that should not be shared with inference.

The model identifier must match. AISRV_LLM_MODEL must exactly match what the endpoint reports under /v1/models; otherwise "Model not found".

No silent fallback. If the endpoint is offline, basebox shows the error and recovers as soon as it is ready again. It does not switch to another model unnoticed – plan monitoring accordingly.

Who decides what

Decision Who
Which models are technically connected Platform Operator (these pages)
Which model the organisation uses by default, which per app Administrator – Model selection · Enabling models
Reasoning effort per chat Users – Reasoning effort
Whether external model providers (OpenAI, Anthropic) are an option The organisation, contractually and under data protection law – Model providers

The models shipped by basebox are listed in the Model Register of the EU AI Act compliance package; basebox modifies models only through quantisation. Whoever retrains or fine-tunes a model assumes the role of provider – see Disclaimer. Closed-source models are not part of the delivery.

Next step: Inference architecture