What can you run?
Sovereignty & privacy
HARDWARE
ILLUSTRATED CONNECTION · HARDWARE FIT NOT VERIFIED
AI MODEL LIBRARY · HUGGING FACE DISCOVERY
Choose the models for your AI system.
Explore Hugging Face model metadata for your hardware and Halo AI workflow. Compare publisher origins, licences and estimated local fit, then validate the selected model within your full operating boundary.
Explore the model library ↓- 1 Choose a task
- 2 Set your hardware
- 3 Review model evidence
SBX AI Hardware supplies local capacity. Halo AI connects your business knowledge, tools and approvals. This library helps select the models; deployment tests establish the actual privacy and control boundary.
Find a model for your work.
Text-generation models; check the exact version for chat or instruction support.
Choose your hardware to establish a target.
Set hardware details
Advanced: system RAM, multiple GPUs & workload
System RAM is not added to VRAM. Multiple GPUs need an explicitly supported sharding strategy. Context and user count are recorded for testing; the estimate below covers weights only, not speed or full runtime memory.
More tasks & search options
Search spans the public Hugging Face catalogue, one page at a time—not only our saved dossiers. Private repositories and unexposed records are outside coverage. Search terms and task filters go through our server to Hugging Face; hardware details stay in your browser.
Loading the public catalogue…
“Memory candidate” is a weight-size screen, not proof it runs. No model has an SBX-tested badge until an exact configuration has passed a recorded benchmark. Lower-bit formats may not be available for every model.
Know what you control.
Origins
Who published it, which models it builds on, and where the evidence stops.
Permissions
Licence terms, access conditions and dependencies to review before use.
Your data
Check offline operation, telemetry and tool access in your deployment.
Country of origin alone is not a sovereignty score. Select a model for its evidence and open questions.
Compare your candidates.
Choose up to three models to compare.
Your model & hardware plan.
Keep the candidates you want to test. Your export includes the selected hardware, model sources and the checks still needed—not an automatic deployment. Each model is screened individually; running several together needs a combined memory and workload test.
No account required. No weights are downloaded and no models are run here.
How sovereignty and compatibility are assessed
We separate model discovery from evidence. Public repository metadata comes from Hugging Face, cached for up to an hour. The eight dated SBX dossiers add publisher references and jurisdiction links. Everything else keeps unknown origins explicit: a repository's language or hosting-region tag is not evidence of its creator's country.
For each deployment, establish the publisher, base model, training-data disclosures, tokenizer, exact weight artifact and packager, licence, runtime, hardware and permitted network routes. Country alone is not a sovereignty score. A memory estimate is not a benchmark or a licence approval.
Nominal weight GB = reported total parameters × selected bits ÷ 8 ÷ 1 billion. We compare with one device's memory after your reserve; we never assume multi-GPU pooling. Sparse, packed and mixed-precision formats require artifact-specific review. KV cache, activations, context and concurrent use require a workload test.
The catalogue can include adapters, quantizations, base checkpoints and non-chat models. Inspect dependencies and official instructions before downloading or executing anything.
Research jurisdictions in the Sovereignty Atlas → · Hugging Face API documentation ↗
Independent SovereigntyBox directory; not affiliated with Hugging Face or model publishers. No compliance certification is implied. The broad catalogue is queried on demand; it is not a complete locally mirrored index.