Model integrations

The models Steward integrates through Phoeniqs MAAS.

Steward uses Phoeniqs Model-as-a-Service for language-model inference, document processing, embeddings, reranking, and transcription. For client-facing chat and drafting work, Steward calls the models directly, without a content-filtering gateway.

Data protection comes from the boundary, not from scanning your words: every request stays inside the Swiss-operated Phoeniqs MAAS environment, and nothing is sent to the model publisher.

Client-facing generation

These are the production defaults for user-facing chat, drafting, analysis, and vision-assisted turns. Steward calls these models directly. Phoeniqs also publishes -GRC aliases that add a content-filtering gateway, but Steward does not use them for client work: their generic safety filters misclassify ordinary family-office topics — portfolio positioning, tax, structuring — as unsafe and block legitimate questions.

Purpose	Model alias	Posture	Notes
General chat and drafting	`inference-llama4-maverick`	Direct	Default for user-facing assistant turns and written work.
Analytical chat	`inference-glm5`	Direct	Default for chart-capable and analysis-heavy chat routes unless overridden.
Reasoning turns	`inference-gpt-oss-120b`	Direct	Used for scenario, risk, comparison, and recommendation questions.
Vision chat	`inference-qwen3-vl-235b`	Direct	Used when a chat turn includes supported image attachments.
Chat fallback	`inference-deepseek-v32`	Direct	Used if the primary model is briefly unavailable, so a turn can still complete.

Specialized processing

These handle document processing, retrieval, and transcription rather than client-facing conversation. Steward routes them through Phoeniqs MAAS inside the same Swiss processing boundary and logs only sanitized operational metadata.

Purpose	Model alias	Posture	Notes
Structured extraction	`inference-qwen3-8b`	Task-specific	Used for JSON-style extraction where deterministic local fallbacks also exist.
Document OCR	`inference-deepseek-ocr`	Task-specific	Used to extract text from scanned or image-heavy documents.
Document parsing	`inference-miner-u25`	Task-specific	Used for document layout and parsing support.
Speech transcription	`inference-whisper-large-v3`	Task-specific	Used for meeting audio transcription when enabled.
Embeddings	`inference-bge-m3`	Retrieval	Used to index and retrieve workspace context.
Reranking	`inference-bge-reranker`	Retrieval	Used to improve retrieval ordering before answer generation.

Evaluation-only models

Steward keeps an admin benchmark path for internal evaluation. That path measures candidate models on behavior, latency, and answer quality before any change to the client-facing defaults above. It is not used to serve client work.

Current benchmark sweep: inference-llama4-maverick, inference-deepseek-v32, inference-qwen3-vl-235b, inference-gemma4-31b, inference-apertus-70b, inference-glm-51-754b, inference-glm45-air-110b, inference-glm5, inference-gpt-oss-120b, inference-llama4-scout-17b.

The model publisher does not receive Steward prompts or outputs. Phoeniqs MAAS serves the models inside the Swiss-operated Phoeniqs environment, and your prompts and their completions stay within that boundary.