Do we need to use Soofi S specifically?

No. We evaluate Soofi S alongside other open-weight and commercial options; the soofi-trainer agent patterns (advisor, RAG, training gateway) apply regardless of which model you land on.

Is Soofi S generally available today?

As of publication, the consortium describes Soofi S as being tested with industry partners, not generally released. We track availability and help you plan around it, including fallback models.

Can we reuse our existing infrastructure?

Usually yes. We integrate with your existing Kubernetes, IaC, and MCP tooling rather than mandating a fresh stack, unless isolation requirements say otherwise.

Soofi training & consulting — LLM specialization on Soofi S

Why Soofi consulting is its own workstream

Soofi S is a 30B Mixture-of-Experts model from a German research consortium, funded under IPCEI-CIS, built for sovereign, EU-hosted industrial AI. The consortium also ships soofi-trainer, an open agent stack (LangGraph advisor, dataset, and training agents; MCP-based vector search and training gateway; RAG on Weaviate) that turns "should we fine-tune or use RAG?" into a guided, auditable process. Getting from evaluation to a production deployment still needs platform engineering, data pipeline work, and a specialization strategy fitted to your use case — that is what we deliver.

Trademark & status notice

Soofi, Soofi S, and soofi-trainer are projects of the Soofi consortium (funded via the IPCEI-CIS programme, funding ID 13IPC040A) and remain their property. We are an independent delivery partner; this page is not an endorsement by, or partnership with, the Soofi consortium. Model access is currently limited to industry testing per the consortium.

What "Soofi training & consulting" covers

LLM specialization advisory

We assess your use case against the same decision space soofi-trainer's advisor agent models — RAG, fine-tuning, LoRA/QLoRA, SFT, DPO — and recommend a method with a defensible cost/benefit case.

RAG & knowledge-base delivery

Markdown/YAML knowledge ingestion, embeddings, and vector search on Weaviate or your preferred store — evaluated against your documents, not a demo corpus.

Training job orchestration

Managed fine-tuning pipelines with a training gateway pattern (MCP-exposed job management), GPU-aware backends, and Grafana/Prometheus observability.

Agent stack integration

LangGraph ReAct agents, A2A orchestration, and AG-UI/SSE streaming for interaction, advisor, dataset, and training agents — wired into your existing tools via MCP.

Delivery services

Evaluation & scoping sprint

A short, structured assessment of your candidate use case, data readiness, and whether Soofi S or an alternative open-weight model fits — with a written recommendation.

Self-hosted deployment

Kubernetes platform engineering and Terraform/OpenTofu-governed environments for model serving, the vector store, and the agent stack — on EU-hosted infrastructure.

Fine-tuning & dataset engineering

Dataset curation, labeling workflows, and training job execution — with the training gateway exposing job status and metrics through MCP tooling.

Quality assurance with Cursor and Claude Code

Fast delivery only pays off when it is controlled and reviewable.

Binding standards

Architecture and coding standards so AI-assisted delivery of agents and pipelines stays consistent and maintainable.

Controlled review processes

Critical changes — model routing, data access, training jobs — are reviewed technically and for security before production.

Documented decisions

Specialization-method choices, data paths, and infrastructure decisions are documented for audit and handover.

Questions on Soofi training & consulting

Do we need to use Soofi S specifically?

No. We evaluate Soofi S alongside other open-weight and commercial options; the soofi-trainer agent patterns (advisor, RAG, training gateway) apply regardless of which model you land on.
Is Soofi S generally available today?

As of publication, the consortium describes Soofi S as being tested with industry partners, not generally released. We track availability and help you plan around it, including fallback models.
Can we reuse our existing infrastructure?

Usually yes. We integrate with your existing Kubernetes, IaC, and MCP tooling rather than mandating a fresh stack, unless isolation requirements say otherwise.

Scope your Soofi S evaluation or training programme

We turn a model announcement into a concrete, reviewable delivery plan — specialization method, infrastructure, and handover included.

Soofi S adoption, LLM specialization, and training delivery

We help you evaluate Soofi S, stand up the soofi-trainer agent stack, and choose the right specialization method — RAG, fine-tuning, or both — on infrastructure you control.