Private AI is more than a model

THOX.ai builds local-first AI systems around the people using them: purpose-built models, personal devices, portable runtimes, private networking, and human-controlled workspaces.

Our goal is practical AI that works closer to your data, on hardware you control. We design for privacy, reliability, and a useful everyday experience—not cloud dependence by default.

This Hugging Face organization is home to our published model checkpoints, adapters, quantizations, model cards, and experimental Spaces. It brings together original THOX model research and clearly identified derivatives of community models.

Run AI. Connect AI. Work with AI.

ThoxBeam™ is the portable AI runtime engine engineered for ThoxKey™, inside ThoxOS—not a separate platform. The drive carries runtime assets, model assets, and workspace resources; a trusted, compatible host supplies execution and working memory.

THOXY™ is the conversational front door. ThoxEmployee provides Digital Employee profiles and lifecycle tooling, while Rolodex introduces the team and its roles. These experiences connect to our broader direction for human-controlled agentic work.

Explore the models

We work across small language models, device-role checkpoints, intent routing, instruction tuning, and deployment-oriented quantization. Different jobs need different models—not every request needs the largest model available.

Start with the THOX edge device models collection.

This is a selected starting point, not the complete catalog. Browse all public repositories for additional model families and formats.

Published weights are not a hardware-performance certificate. Model names identify families or intended roles; they do not establish that a model runs on every similarly named device. Check the exact model, artifact, runtime, hardware revision, and published validation evidence.

Choose the right runtime

GGUF / llama.cpp / LM Studio: Use a published GGUF artifact with a runtime that supports its architecture and quantization. A .gguf extension alone does not guarantee compatibility.

Ollama: Use the exact repository and tag published on our Ollama profile. Hugging Face and Ollama names, formats, and availability can differ.

Transformers / PEFT: Follow the individual model card. Standard checkpoints, custom architectures, and LoRA adapters have different loading requirements. An adapter is not a standalone replacement for its base model.

ThoxOS / ThoxBeam / MeshStack: Match the model to a supported execution path—device-local, compatible-host, or authorized peer. Storing a model on a device is not the same as running inference on that device.

For deployment guidance, see the model catalog and compatibility documentation alongside each repository's release notes.

Local first. Explicit connections. Human control.

Our design principles are local-first execution, permissioned access to private resources, explicit authorization for external processing, and reviewable agent actions. Model output is not authorization. A routing suggestion or agent plan should never grant itself access to your files, tools, or devices.

Offline operation depends on the selected model, runtime, locally available assets, and workflow dependencies. Hosted Spaces and online demos may process data remotely; consult their disclosures before submitting sensitive information. A hosted demonstration is not proof of offline execution.

Meet the devices

The current four-device lineup is ThoxKey™, ThoxAir™, ThoxMini™, and ThoxClip™.

Device support and availability are release-specific. Historical ThoxNova model names do not announce shipping ThoxNova hardware; ThoxNova remains a roadmap device.

See the THOX website and campaign page for current product and campaign information.

Licensing and provenance

Licensing is per repository and per artifact. Read the relevant LICENSE, model card, and upstream terms before use or redistribution. Do not infer a license from a model family, a THOX name, or public download availability.

Some device-role repositories carry licensing marked for review. Treat those terms as unresolved until the repository provides clear permission for your intended use.

The apache-2.0 metadata on this organization-card Space is not a blanket license for every THOX model, derivative, runtime, or product. THOX-developed artifacts and third-party components remain subject to their respective published terms.

Build with us

We welcome reproducible evaluations, compatibility reports, documentation improvements, and thoughtful collaboration on efficient, user-controlled AI.

For a model issue, use that repository's Community tab and include the model revision, artifact or quantization, runtime version, hardware, and a minimal reproduction. Never include credentials or private user data.

Training tooling · Device/model mapping · Documentation · THOX.ai

THOX.ai LLC — Your AI. Your Data. Your Rules.™

Organization overview updated October 4, 2026. Individual model cards and release notes describe artifact-specific status, requirements, and limitations.