Preview · August 2026
German Whitepaper
Complete German-language presentation of the platform, architecture and research vision.
- Length
- 118 pages
- Format
- PDF · 3.1 MB
Open source · Apache 2.0 · Self-hosted
The AI infrastructure that turns prompts into task-tailored workflows.
MoE Sovereign analyses each request, combines the right models, Expert Templates, GraphRAG, MCP tools and current research—then validates the result and retains useful knowledge for future tasks.
Not just a router. A compiler for executable AI workflows.
- Local inference
- Familiar APIs
- Versioned Privacy-by-Design self-assessment
Prompt 01
Expert TemplateGraphRAG
MCP ToolsWeb Research
Meta-OrchestratorMoESovereignAnalyse · Plan · Route
LLMs / SLMsagentic_tool
Judge + ValidationKnowledge + Memory ↺
Answer 06
One API, many capabilities
One LLM to clients.
A full system inside.
OpenAI Chat CompletionsOpenAI ResponsesAnthropic MessagesOllama API
POST /v1/responses
{
"model": "moe-sovereign",
"input": "Assess the current situation …"
}Whitepaper · Preview Release
Architecture, research and evidence—open for review.
All three documents are preview editions. Content, methodology, results and pagination may change before final publication. Feedback is explicitly welcome.
Preview · August 2026
English Whitepaper
Complete international edition covering the platform, architecture and research vision.
- Length
- 118 pages
- Format
- PDF · 3.1 MB
Preprint · Preview
Scientific short paper
Concise English research edition for scientific discussion and review.
- Length
- 13 pages
- Format
- PDF · 355 KB
Publication status: Preview / preprint. Not yet released as a final edition and—unless stated otherwise in the document—not peer-reviewed.
Same prompt. Different process.
A model answers. MoE Sovereign organises the solution.
The architecture spends compute where freshness, knowledge, multiple steps or evidence justify it.
Direct model operation
Native LLM
- PromptOne input
- One modelStatic training knowledge
- AnswerNo persistent learning experience
For simple, stable tasks, this direct path may be optimal.
Adaptive processing
MoE Sovereign
- Prompt analysisDomain, complexity, risk
- Expert TemplateModels, dependencies, policies
- ExecutionGraphRAG · MCP · research · agents
- ValidationJudge · sources · correction loop
- Answer + knowledgeUseful results persist
The greatest advantage emerges for complex, current, knowledge-intensive or safety-relevant tasks.
The learning architecture
Knowledge grows outside model weights.
Today, a large model solves a novel task. Tomorrow, a smaller model can reuse verified knowledge and a proven workflow.
01ProcessExecute workflow
→02ValidateCheck quality
→03ExtractFacts, sources, paths
→04PersistGraph + memory
→05ImproveRouting + next request
Production system learning
GraphRAG, layered caches, episodic knowledge, Correction Memory, feedback and routing signals improve system context—without uncontrolled training.
Formal verification & quality gates
Research and development of deterministic GBNF constraints, Kahn DAG cycle checks, SMT-based validation and entropy-based egress guards.
Model distillation on LUMI-G
Knowledge from large models and verified plans are transferred on EuroHPC LUMI-G into specialized small models for resource-efficient execution.
One sovereign ecosystem
Three projects. Three clearly separated responsibilities.
MoE Sovereign remains the independently deployable core. Libris and Codex optionally extend it with federated knowledge exchange and an auditable data platform.
S
Core · Orchestration
MoE Sovereign
Compiles prompts into workflows of models, Expert Templates, knowledge, tools and validation. Operates fully without its sibling projects.
- Adaptive AI workflows
- GraphRAG, MCP and memory
- OpenAI, Anthropic and Ollama APIs
L
Federation · Knowledge
MoE Libris
Federated knowledge exchange between sovereign instances—voluntary, trust-based and reviewed before acceptance.
- JSON-LD push and pull
- Pre-audit and human review
- mTLS and sender-side privacy scrubbing
C
Data · Audit
MoE Codex
Optional sovereign data and audit platform for organisations requiring catalogues, approvals, lineage and evidence.
- Data catalogue and versioning
- Approvals, lineage and drift detection
- Policies, search and analytics tools
Modular, not monolithic: Each project has its own responsibility and can be developed or operated independently. Integrations use documented interfaces.
Quality through interaction
Not more features. Structured execution & validation.
Better knowledge
GraphRAG, documents and current research provide grounded context beyond training knowledge.
Precise execution
MCP precision tools and structured tool paths execute deterministic tasks reliably.
Better decisions
Expert Templates, complexity analysis and routing select only necessary resources.
Higher reliability
Judge, Self-Correction, contradiction detection and source ranking check before output.
Continuous improvement
Corrections, feedback, episodes and routing scores return as reusable system context.
Reference architecture
Six layers. One familiar API outside.
01InterfaceOpenAI · Anthropic · Ollama · existing clients
02IntelligenceMeta-orchestrator · planner · templates · routing · policies
03ExecutionLLMs · SLMs · expert workers · agentic_tool · MCP · research
04KnowledgeNeo4j GraphRAG · vectors · documents · cache · memory
05QualityJudge · Self-Correction · Corrective RAG · security gates
06ObservabilityTracing · metrics · audit · routing telemetry
EuroHPC · LUMI-G
Use supercomputers so the platform needs fewer supercomputers later.
One-time intensive compute supports research into planning, routing and validation. The deployment target is smaller local specialist models—not permanent HPC dependence.
Support is provided by the EuroHPC Joint Undertaking as an allocation of high-performance computing resources, not as a direct financial grant.
View research documentation →ImplementedCompound-AI pipelineTemplates, GraphRAG, tools, validation, memory
In validationTraining and evaluation dataTraceable plans and quality decisions
In developmentLUMI-G distillationComplexity, routing, tool selection, validation
PlannedSpecialised SLMsMore local handling of recurring tasks
Research corner
Test hypotheses—not just demonstrate systems.
The research asks which parts of a compound-AI system can measurably learn, when persistent knowledge can offset model size, and how quality can be assessed transparently.
Q₀
Can a learning orchestration and knowledge architecture enable small, locally deployable models to handle demanding tasks with improving quality and more efficient repeat processing?
R1
In developmentRouting & plan distillation
Hypothesis: Successful expert plans can be transferred into smaller routers and planners without asking a large model to re-plan every request.
- Method
- Supervised fine-tuning, routing evaluation and ablations on LUMI-G
- Measures
- Plan quality, routing accuracy, latency and resource use
R2
In validationTrust metrics & Judge
Hypothesis: Source, graph and contradiction signals can support more reliable quality decisions than model self-assessment alone.
- Method
- Paraconsistent conflict assessment, trust scores and controlled comparisons
- Measures
- Factuality, conflict detection, calibration and error rates
R3
System basis activeKnowledge accumulation
Hypothesis: Verified facts, solution paths and corrections shorten similar future processing and reduce repeated model work.
- Method
- GraphRAG, episodic memory, semantic caches and repeated task series
- Measures
- Cache/knowledge hits, latency, tokens, reuse and answer quality
Methodological guardrails
Reproducible, status-clear, falsifiable.
- Document configuration, version and hardware
- Separate internal measures from external baselines
- Report sample, variance and limitations
- Do not hide negative or failed results
From prompt to retained knowledge
Concrete workflows, not industry labels.
Public administrationReview procurement documents against current rules
“Check this specification for contradictions and missing evidence.”
- Template
- Document analysis → legal sources → contradiction check
- Resources
- Local LLM, documents, GraphRAG, current official sources
- Validation
- Source rank, citations, Judge, uncertainty labels
- Retained
- Approved terms, relations, corrections and review patterns
IT & securityAssess an architecture against threats
“Build a threat model for this hybrid platform and prioritise controls.”
- Template
- System model → threat research → controls → review
- Resources
- Security expert, MCP, internal standards, current advisories
- Validation
- Coverage, source recency, contradictions, Judge
- Retained
- System relations, accepted risks and validated controls
Software engineeringSolve a reproducible production fault
“Find the cause of this intermittent failure and provide a tested patch.”
- Template
- Reproduce → analyse code → patch → test → review
- Resources
- Code SLM/LLM, repository tools, tests, documentation
- Validation
- Build, tests, static analysis, diff review
- Retained
- Root cause, verified solution path and relevant code relations
Evidence, not superlatives
Metrics with context.
Internal benchmarks describe the real behavior of a concrete configuration, separating model capabilities from system architecture.
Deterministic VLSM precision
Typed MCP computation produced correct subnet partitioning in the VLSM task; ungrounded generation produced overlapping ranges in 80% of comparison cases.
Observed within the August 2026 pre-distillation benchmarkMethod and artefacts ↗Temporal consistency
In a controlled policy-change task, GraphRAG used the current Port 8443/mTLS state while the native baseline reverted to obsolete training knowledge. Result reproduced across two rounds.
Controlled task comparison · August 2026Inspect benchmark ↗Observed latency difference
Across five internal epochs. Graph growth, cache reuse, correction memory and model warmth changed simultaneously; individual causal contributions were not isolated.
Longitudinal observation · 2026Ablation data ↗Sovereignty & security
Control is a technical property—not a blanket legal claim.
MoE Sovereign provides technical prerequisites for controllable data flows and data protection by design. Legal assessment depends on the actual deployment, purpose and organisational controls.
Review the versioned privacy self-assessment →- Local inference and optional offline/air-gap operation
- Explicitly configurable outbound connections
- Deterministic execution plans with Kahn DAG cycle verification
- Research & development: Deterministic GBNF constraints, SMT solvers and entropy guards
- Versioned Expert Templates and configuration
- Transparent audit and telemetry persistence in PostgreSQL for robust traceability
- OCI-portable, containerised and rootless operation
- Security gates and hardened tool paths
Deploy to fit
From local system to platform.
An OCI-based architecture supports different operating models. Kubernetes and OpenShift are architecturally prepared; production validation remains deployment-specific.
ONE-LINE INSTALLER · LINUX
Install MoE Sovereign interactively
For a fresh Debian or Ubuntu server. The installer detects Docker or Podman, generates secrets and guides the configuration.
curl -sSL https://raw.githubusercontent.com/h3rb3rn/moe-sovereign/main/install.sh | bashReview shell scripts before execution: inspect install.sh · requirements and manual installation
Notebook & Edge (WSL2)
16/32 GB RAM standard laptops with CPU-optimised inference and high processing speed.
GPU Workstation
Multi-GPU setups (NVIDIA/AMD) for large context windows and parallel expert execution.
Enterprise Cluster
Scalable platform (AIHUB, Ollama, Kubernetes) for multi-tenant high throughput and auditing.
LXC · Docker Compose · Podman Quadlet · Kubernetes · OpenShift
The architecture is open