ai_provider_mittwald 1.2.0

This release adds two new operation types (reranking and text-to-speech) and brings the model lineup in sync with what mittwald AI Hosting currently offers.

New features

• Reranking is now supported, for improving RAG retrieval precision by reranking search results. Default model: Qwen3-VL-Reranker-2B (text and image scoring).
• Text to Speech is now supported (#3620140). Default model: Qwen3-TTS-12Hz-1.7B-CustomVoice, with nine voices (aiden, dylan, eric, ono_anna, ryan, serena, sohee, uncle_fu, vivian), configurable speaking rate and spoken language, and mp3, opus, flac, wav or pcm output. Voice cloning is not available.
• New chat model: Qwen3.8-27B-NVFP4, with reasoning and image vision (#3620139).

Changes to the model lineup

Mistral Small, Qwen3 Coder and Devstral have been withdrawn by mittwald and are no longer offered by this module. Sites that had one of them configured are migrated to Ministral-3-14B-Instruct-2512 automatically — see the upgrade notes below.

Qwen3.5-0.8B is no longer offered for image input. It is a text-only model and was previously matched by an over-broad vision filter.

Upgrade notes

• This release requires drupal/ai:^1.3.0 (previously ^1.2.0), for the reranking operation type interface. Update the AI module first.
• Run database updates (drush updb, or /update.php) after updating. Two update hooks migrate any configured model that mittwald no longer offers. Note that sites which already ran update_10001 on 1.1.0 still need update_10002: the earlier hook had an inverted provider check and silently skipped exactly the providers it was meant to migrate.