Model Registry
Not every model qualifies for sovereign AI.
Sovereignty is not a model feature — it is a deployment decision. Self-hosting provides control options; sovereignty claims require evidence from the actual deployment. This registry maps the distinction clearly.
Sovereignty Tiers
Three tiers. One question: who controls the inference?
Sovereignty is determined by where inference happens, who owns the runtime, and whether your data crosses a boundary you do not control.
Self-hosted candidate
Model runs entirely on infrastructure you control. Validate external calls, telemetry, licenses and dependencies before claiming that data stays within the approved boundary.
- Open-weight model you can download and self-host
- Runs on your own hardware, VMs, or private cloud
- Zero external API calls during inference
- License, artifact provenance, runtime and data controls are verified
Conditionally Sovereign
Managed model services retain material provider dependencies. Evaluate the exact service and deployment configuration against your required data and operating boundary.
- Account access controls and processing locations verified separately
- Verify processing location, retention and provider access
- Provider controls model weights and updates
- Dependency on provider availability and pricing
Not Sovereign
Inference happens on the provider's shared infrastructure. Your prompts, documents, and responses leave your boundary and are processed externally.
- API calls route to provider's shared servers
- Processing location and access require provider due diligence
- Subject to provider's data retention policies
- Data classification and approved terms determine acceptable use
Self-hostable model candidates
Model candidates for a controlled deployment
Open weights permit self-hosting subject to their licenses. Actual sovereignty depends on runtime identity, telemetry, egress, dependencies, updates and recovery; model family alone establishes none of those controls.
*Candidate fit is editorial orientation, not a benchmark or approval. Latest catalog additions reviewed on 27 October 2026; each sourced entry has its own review date. Older entries remain comparison baselines. Runtime and quantization compatibility must be verified against a pinned artifact and build. Model-specific sizing candidates are excluded from automatic matchmaker fit results.
Evaluated candidates: none with published workload measurements yet. The shortlist below is based on source review, not performance validation.
Reviewed candidates: evaluation pending
| Model | Provider | Deployment | Context | Candidate fit* | Use cases |
|---|---|---|---|---|---|
gpt-oss-20b Apache-2.0. MoE with mixed sliding/full attention. MXFP4 expert weights are not interchangeable with a generic Q4 estimate; verify backend support, cache allocation and Harmony formatting. Model card: reviewed 2026-10-27 | OpenAI | MXFP4-compatible runtime or verified converted artifact Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 128K configured | Medium | ReasoningMoESizing required |
gpt-oss-120b Apache-2.0. Budget all resident experts, mixed-attention cache and runtime overhead. Publisher memory examples are not guarantees at maximum context or concurrency; benchmark the exact artifact. Model card: reviewed 2026-10-27 | OpenAI | MXFP4-compatible runtime or verified converted artifact Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 128K configured | Medium | ReasoningMoESizing required |
Qwen3.8-27B Apache-2.0. Hybrid DeltaNet/attention with a vision encoder. Extended context requires explicit configuration. Benchmark cache, recurrent state and image memory for the chosen artifact. Model card: reviewed 2026-10-27 | Qwen | Hybrid multimodal runtime; model-specific sizing required Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 256K native | Medium | MultimodalHybridSizing required |
Gemma 4 31B Apache-2.0. Dense multimodal model. Mixed attention and image inputs require cache and encoder budgeting beyond a parameter-only estimate. Model card: reviewed 2026-10-27 | Google DeepMind | Gemma 4 compatible runtime; model-specific sizing required Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 256K | Medium | MultimodalDenseSizing required |
Gemma 4 26B A4B Apache-2.0. Budget resident weights for the full MoE, not only active parameters. Validate attention cache and multimodal overhead on the selected runtime. Model card: reviewed 2026-10-27 | Google DeepMind | Gemma 4 MoE compatible runtime; model-specific sizing required Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 256K | Medium | MultimodalMoESizing required |
Nemotron 3.5 Bolt E4B 4GB OpenMDW-1.1. 8B total / 4B effective parameters. The 4GB artifact label is not a total serving-memory guarantee: add cache, runtime and any speculative drafter. Model card: reviewed 2026-10-27 | NVIDIA | Supported GGUF runtime; model-specific sizing required Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 64K | Medium | EdgeQuantizedSizing required |
Nemotron 3.5 Bolt 74B A9B 24GB OpenMDW-1.1. Quantized MoE. The 24GB label is not a fit guarantee at arbitrary context or concurrency. Verify IQ2 kernel support and the exact runtime/hardware requirements. Model card: reviewed 2026-10-27 | NVIDIA | Artifact-supported runtime and GPU; model-specific sizing required Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | Up to 256K | Medium | MoEQuantizedSizing required |
DeepSeek-V4.1-Flash-1017 MIT. Asymmetric encoder-decoder architecture and compressed cache need dedicated sizing. Active parameter counts do not represent resident weight memory. Validate long-context capacity separately. Model card: reviewed 2026-10-27 | DeepSeek | Distributed model-specific serving stack Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 1M published | Medium | MoELong contextSizing required |
Qwen3-8B Apache-2.0. 8.2B dense parameters. Extended context needs explicit configuration and additional memory; benchmark the selected artifact. Model card: reviewed 2026-10-20 | Qwen (open weights) | Compatible local runtime; pin release Artifact revision: b968826d9c46dd6066d109eabc6255188de91218 Runtime build: vllm/vllm-openai:v0.11.2 (generator target; unbenchmarked) Measured results: not published | 32K native | Medium | MultilingualAgents |
Qwen3-Coder-30B-A3B-Instruct Apache-2.0. MoE: 30.5B total, 3.3B active parameters. Plan weight memory for all experts, then budget cache and serving overhead. Not a drop-in replacement for the basic lab model. Model card: reviewed 2026-10-20 | Qwen (open weights) | Compatible local coding runtime; pin release Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 256K native | Medium | CodeAgents |
Qwen3.5-9B Apache-2.0. Hybrid architecture and multimodal inputs need model-specific state and encoder memory planning. Not included in the parameter-only matchmaker shortlist. Model card: reviewed 2026-10-20 | Qwen (open weights) | Model-card-supported runtime; validate exact build Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | See pinned model configuration | Medium | MultimodalHybrid |
Historical baselines
| Model | Provider | Deployment | Context | Candidate fit* | Use cases |
|---|---|---|---|---|---|
Llama 3.1 8B / 70B / 405B Established comparison baseline. Check the exact variant, license and current runtime support before selection. Historical baseline: review date unavailable; source review required | Meta (open weights) | Ollama · vLLM · OpenShift AI Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 128K | High | RAGAgentsSummarization |
Mistral 7B / Mixtral 8×7B Older dense and MoE comparison baselines. Measure quality, total expert memory and latency on your workload. Historical baseline: review date unavailable; source review required | Mistral AI (open weights) | Ollama · vLLM · Docker Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 32K | High | RAGClassificationCode |
Mistral Nemo 12B Older 12B comparison baseline. Validate long-input recall and memory rather than inferring quality from context length. Historical baseline: review date unavailable; source review required | Mistral AI (open weights) | Ollama · vLLM Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 128K | High | DocumentsRAGLong context |
Phi-3 / Phi-3.5 Mini Small-model comparison baselines. Verify the exact variant and measure device memory and task quality. Historical baseline: review date unavailable; source review required | Microsoft (open weights) | Ollama · Edge · Docker Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 128K | Medium | EdgeLightweightClassification |
Gemma 2 9B / 27B Older comparison family. Validate task quality and license terms for the selected artifact. Historical baseline: review date unavailable; source review required | Google (open weights) | Ollama · vLLM Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 8K | Medium | RAGReasoningSummarization |
Qwen 2.5 7B / 72B Retained comparison family. Evaluate each language and task; the basic coding lab pins a separate qwen2.5-coder:7b artifact. Historical baseline: review date unavailable; source review required | Alibaba (open weights) | Ollama · vLLM Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 128K | High | MultilingualCodeRAG |
DeepSeek R1 / V3 Reasoning and general-purpose variants have different configurations. Verify exact artifact, license, context and full serving memory. Historical baseline: review date unavailable; source review required | DeepSeek (open weights) | vLLM · OpenShift AI Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 64K | High | ReasoningAnalysisCode |
Code Llama / StarCoder2 Purpose-built for code generation, completion, and review in private developer tooling. Historical baseline: review date unavailable; source review required | Meta / BigCode (open weights) | Ollama · vLLM Artifact revision: Not pinned Runtime build: Not validated Measured results: not published | 16K | Medium | CodeDeveloperCompletion |
Conditionally Sovereign Models
Cloud-managed examples requiring deployment review
Managed offerings require version, region, retention, support-access and subprocessor review. A tenant or private endpoint alone does not establish exclusive data control.
Conditional sovereignty depends on your cloud agreement, region selection, and data processing terms. Always verify provider data residency commitments before using these models with sensitive data.
| Model | Provider / Platform | Deployment | Context | Candidate fit* | Use cases |
|---|---|---|---|---|---|
Mistral Large 4 (preview) The 6 October announcement describes an API preview and planned weight release. Do not assume downloadable weights, licensing or production availability from that announcement. Official source: reviewed 2026-10-27 | Mistral | API preview; self-hosted release not verified | Verify preview endpoint limits | Medium | PreviewMultimodal |
GPT-4o / GPT-4 Turbo Historical managed-model example. Verify current catalog availability, processing geography, retention, access and contractual controls. | Microsoft via Azure OpenAI | Azure OpenAI Service | 128K | High | RAGAgentsEnterprise |
GPT-3.5 Turbo Legacy managed-model example; check availability, price, region and workload quality before considering it. | Microsoft via Azure OpenAI | Azure OpenAI Service | 16K | High | RAGSummarizationHigh volume |
Claude 3.x (Haiku / Sonnet / Opus) Legacy managed-model example. Verify current model version, supported context, processing terms and access controls. | Anthropic via AWS Bedrock | AWS Bedrock | 200K | High | DocumentsLong contextRAG |
Llama 3 (Managed) Legacy managed deployment example. Review provider control, license, model version and exit options. | Meta via AWS Bedrock / Azure | AWS Bedrock · Azure AI | 128K | High | RAGManagedTransition |
Mistral Large / Small Legacy managed deployment example. Confirm provider, region, data controls and currently available model versions. | Mistral via Azure AI | Azure AI Foundry | 32K | Medium | RAGManagedAzure |
Not Sovereign — Public APIs
Models that cross your data boundary
Direct managed APIs create an external processing dependency. Check current model availability, terms, retention and processing controls before sending data; these historical entries are not a current provider catalog.
Public API models are listed here for comparison, not recommendation. For many use cases without sensitive data, public APIs are practical. For sovereign AI, they are excluded by definition.
| Model | Provider | Deployment | Context | Sovereign fit | Sovereign alternative |
|---|---|---|---|---|---|
GPT-4o / GPT-4 Legacy direct-API example; external processing must be assessed against data classification and current contractual controls. | OpenAI (direct API) | OpenAI API | 128K | Not Sovereign | Azure OpenAI Service |
Claude 3.x Legacy direct-API example. Review processing, retention, identity, residency and operational dependencies for the selected service. | Anthropic (direct API) | Anthropic API | 200K | Not Sovereign | AWS Bedrock |
Gemini 1.5 Pro / Flash Legacy direct-API example; verify current service terms and whether its processing controls meet the required boundary. | Google (direct API) | Google AI Studio / API | 1M | Not Sovereign | Vertex AI (GCP) |
Deployment Runtimes
Where and how sovereign models run
The runtime determines the sovereignty level, performance envelope, and operational complexity. Choose based on your infrastructure maturity and compliance requirements.
Ollama
Local and single-server model runtime. Ideal for development, proof-of-concept, and low-volume private deployments.
- Best for
- Development · Laptops · Small teams
- Models
- Verify exact model and quantized artifact support
- Note
- One-command model pull and serve. CPU execution may be feasible for smaller artifacts; validate latency and concurrency.
vLLM
High-throughput inference server with continuous batching. Purpose-built for GPU-accelerated production workloads.
- Best for
- Production · GPU servers · High concurrency
- Models
- Verify exact architecture, kernels and tool-call parser
- Note
- OpenAI-compatible API surface. Supports tensor parallelism for large models. Preferred for production private RAG.
OpenShift AI
Red Hat's ML platform for enterprise Kubernetes deployments. Integrates model serving with observability and governance.
- Best for
- Enterprise · Kubernetes · Air-gapped
- Models
- Verify model architecture against your serving image and accelerator
- Note
- Preferred for regulated industries, air-gapped environments, and enterprises already using OpenShift.
Azure OpenAI Service
Managed model endpoints; confirm supported model versions, deployment type, region and contractual controls.
- Best for
- Azure-aligned orgs · Compliance · Fast start
- Models
- Check the provider catalog for your region and deployment type
- Note
- No self-managed GPU infrastructure required. Verify processing region, retention, provider access, version behavior and deployment terms.
AWS Bedrock
Managed foundation model API authorized through AWS IAM. Verify model and inference-profile routing separately from private network connectivity.
- Best for
- AWS-aligned orgs · Multi-model · Compliance
- Models
- Check the current regional catalog and inference profile
- Note
- Processing location depends on the model, endpoint and inference profile. India geographic cross-Region inference can route between Mumbai and Hyderabad. Review destination Regions, IAM/SCP restrictions, logging and retention. PrivateLink controls network connectivity; it does not guarantee single-Region processing.
Decision Guide
Which model path fits your situation?
Sovereignty requirements, infrastructure maturity, and compliance obligations determine the right deployment path — not model quality rankings.
Situation
You handle regulated data (HIPAA, GDPR, financial)
Recommendation
Approved self-hosted or managed deployment with verified controls
Self-host on vLLM / OpenShift AI, or assess Azure OpenAI / AWS Bedrock with verified processing destinations and access controls
Situation
You have a GPU-equipped private server or Kubernetes cluster
Recommendation
Self-hosted candidate — self-hosted
Evaluate the reviewed self-hostable candidates below with a compatible runtime. Pin weights and serving images, then measure your actual workload before sizing a deployment.
Situation
You want private AI but don't have GPU infrastructure yet
Recommendation
Conditionally Sovereign — managed cloud
Azure OpenAI (if Azure-aligned) or AWS Bedrock (if AWS-aligned). Plan migration to self-hosted as GPU capacity grows.
Situation
You need to run AI on laptops or edge devices
Recommendation
Self-hosted candidate — local runtime
Evaluate a compact quantized candidate with a supported local runtime. Budget shared memory and verify offline operation, latency and application dependencies on the target device.
Situation
You need maximum model quality right now
Recommendation
Conditionally Sovereign
Compare current provider endpoints and self-hosted candidates on your evaluation set. Verify processing destinations, retention and provider access for the exact deployment configuration.
Situation
You are building a prototype or internal demo
Recommendation
Start with Ollama locally, plan for vLLM in production
Choose a reviewed candidate supported by your local runtime. Design the application layer to swap the model endpoint without rewriting the app; evaluate before production use.
Registry Principles
How SovAIHub evaluates models
This registry does not rank models on benchmark performance. It evaluates them on sovereignty, deployment control, and enterprise operational fit.
Boundary control
Where does inference happen? Who owns the compute? Can you prevent data from leaving your network? These questions determine sovereignty, not model size or capability.
Deployment operability
A sovereign model you cannot realistically operate is not a useful recommendation. Models are rated on whether enterprise teams can deploy, monitor, and maintain them.
Weight availability
Fully sovereign models require open or licensed weights you can download. A model that requires an external API for every inference is not self-hosted, regardless of marketing language.
Enterprise context fit
Context window, throughput, and accuracy on enterprise tasks (document Q&A, summarization, classification) matter more than general benchmark scores for private RAG workloads.
Regulatory alignment
Models deployed in regulated industries must support data residency, audit logging, and access control. The registry notes which deployment methods support these requirements.
Runtime independence
SovAIHub favors models that can be served across multiple runtimes (Ollama, vLLM, OpenShift AI) without vendor lock-in. Application code should route to a model endpoint, not a provider.
Next Step
Need help selecting and deploying the right model stack?
Model selection depends on your data classification, infrastructure, compliance requirements, and team capability. SovAIHub can help you map the right path.
Select the right model. Deploy it on infrastructure you control.
Share your use case, data sensitivity, infrastructure environment, and compliance requirements. We will help you identify a practical sovereign model deployment path.