THOX.ai LLC

Back THOX on Kickstarter →

Your AI. Your Data. Your Rules.™ — THOX builds a private AI ecosystem: four devices, the models that run on them, and the mesh that connects them. Everything on this page is public and ready to run.

109 public models in this organization · 89 models on Ollama · 93-entry Model Factory catalog, 91 published

Model Factory · Flagship models · Kernel family · Devices · Browse by category · Licensing

Start here

Our flagship device models

Four models, one per THOX device, trained in the THOX Wave C flagship program and scored on GATE49, our 49-question factuality gate (verified facts, careful declines, known-false traps). Full method and results: the Wave C whitepaper in the THOX whitepapers dataset.

The models run on a host computer today. Device builds are qualified on physical hardware as the devices ship; each card states exactly what has been measured.

THOX Model Factory — find the right model for your machine

Open the THOX Model Factory →

Pick your machine or THOX device, filter by memory and workload, compare the evidence on each card, and hand the pick to your agent. No device scan — you control the information you enter. The catalog carries 93 entries, 91 of them published models, and every published model is here and on Ollama.

Walkthrough (3:47): the THOX Model Factory guided tour — choose your hardware, understand the models, compare the evidence.

ThoxOS Kernel family

New on November 1, 2026: THOX-trained models for ThoxOS kernel and operating-system work, released as open weights.

Research: the ThoxOS research draft — doi.org/10.5281/zenodo.23876230.

ollama run Thox-ai/thoxos-kernel-27b
ollama run Thox-ai/thoxos-kernel-0.8b

Run a flagship now

GGUF builds are served on Ollama:

ollama run Thox-ai/thoxair-0.5b
ollama run Thox-ai/thoxedge-0.8b
ollama run Thox-ai/thoxmini-phi4mini
ollama run Thox-ai/thoxclip-qwen3-0.6b

Or hand one to your agent:

ollama launch claude  --model Thox-ai/thoxair-0.5b
ollama launch hermes  --model Thox-ai/thoxair-0.5b

Notes that matter: ThoxEdge-0.8B works best with thinking disabled ("think": false). ThoxClip-Qwen3-0.6B must be served with its grounding system prompt — its card explains why. Weights on this page load with Transformers; check each card for its exact files and license.

Our Devices

ThoxKey — Your Private AI, from One Computer to the Next.	ThoxAir — Your THOX Workspace, Wireless and on the Go.
ThoxMini — Local Edge AI That Fits in Your Hand.	ThoxClip — Charge Your Phone. Carry Your AI.

For more info: thox.ai.

Our Platforms

Also in the catalog

• Device models: ThoxEdge-0.5B, ThoxFlux-350M, ThoxNano-135M, ThoxNano-360M — earlier-wave models for the same device slots, each with its gate record on its card.
• Device and edge, new at 4B: ThoxEdge-4B (15.19 tok/s) and ThoxClip-Agent-4B (15.28 tok/s) — THOX-configured Qwen3-4B builds, not THOX-trained; 3.89 GB footprint, Knowledge Pack grounding 8/8, fabrication trap declined. ollama run Thox-ai/thoxedge-4b · ollama run Thox-ai/thoxclip-agent-4b
• ThoxWork: ThoxWork-Office-4B (15.32 tok/s, grounding 7/8), ThoxWork-Mind-4B (15.32 tok/s, grounding 8/8), ThoxWork-Analyst-8B (8.73 tok/s, grounding 8/8) and ThoxWork-Agent-8B (8.59 tok/s, grounding 8/8) — THOX-configured Qwen3 builds for the agentic workspace, not THOX-trained; 3.89 GB (4B) and 6.68 GB (8B), fabrication traps declined.
• iOS and device knowledge: ThoxIOS-Open-4B (15.34 tok/s, grounding 8/8, 3.89 GB) and ThoxIOS-Mesh-8B (8.38 tok/s, grounding 7/8, 6.68 GB) — THOX-configured Qwen3 builds for the iOS and device tier, not THOX-trained; fabrication traps declined.
• MeshStack groups: ThoxMeshGroup-Federated-8B — THOX-configured Qwen3-8B build for federated MeshStack workgroups, not THOX-trained; 8.75 tok/s, 6.68 GB, grounding 8/8, trap declined. MeshStack routes whole requests to a node; it does not split inference. ollama run Thox-ai/thoxmeshgroup-federated-8b
• Agents and routing: ThoxAgent-Core-4B — THOX-configured Qwen3-4B agent build, not THOX-trained; 15.12 tok/s, 3.89 GB, grounding 8/8, trap declined. ollama run Thox-ai/thoxagent-core-4b
• Research: ThoxLLM and ThoxMicro are THOX from-scratch model research; the -role checkpoints are small-model role research. Interesting, documented, and not shippable assistants.
• Adapters: repos marked as adapters require their base model and are not standalone downloads.

Start from the THOX edge device models collection rather than the full repo list.

Browse the catalog by category

Plus the earlier role and micro models in the THOX edge device models collection.

Licensing

Licensing is per repository and per artifact. Read the model card and LICENSE in the repo you download; do not infer a license from a model family or from public availability. Most flagship derivatives inherit permissive base licenses (Apache-2.0 or MIT) — the card on each repo is the authority. THOX models default to Apache-2.0 unless a base model or other terms say otherwise; where they do, the card says so plainly.

Build with us

Model issue? Use that repo's Community tab with the model revision, file or quantization, runtime version, and a minimal reproduction. Never include credentials or private data.

Also public

Four more THOX models are now fully public (ungated) on Hugging Face and Ollama:

• ThoxForge-7B — code-oriented 7B assistant, LoRA adapter + Q4_K_M GGUF. ollama run Thox-ai/thoxforge-7b
• ThoxGlobal-7B — multilingual 7B assistant, LoRA adapter + Q4_K_M GGUF. ollama run Thox-ai/thoxglobal-7b
• ThoxNova-12B-Core — conservative chat 12B, LoRA adapter + Q4_K_M / Q5_K_M GGUF. ollama run Thox-ai/thoxnova-12b-core
• ThoxNova-12B-Agent — tool-calling agent 12B, LoRA adapter + Q4_K_M / Q5_K_M GGUF. ollama run Thox-ai/thoxnova-12b-agent

Community redistributions: safety notice

Two repos in this organization redistribute community uncensored models. They are not trained by THOX and are not part of the ThoxOS or MeshStack catalog:

• ThoxHeretic-27B is fully public and ungated. It is an uncensored (abliterated) redistribution, not safety-tuned, with reduced refusal behaviour; apply your own safety layer before any user-facing use.

• ThoxMythos-9B is a research uncensored redistribution, fully public and ungated, with its safety notice on its card; apply your own safety layer before any user-facing use.

• ThoxOpen-8B-Heretic is THOX's own abliteration build, not a redistribution: refusals 13/100 against 100/100 stock in THOX's standalone re-evaluation, KL divergence 0.0482. Its card carries the full method and safety notice.

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

Organization overview updated November 2, 2026. Model cards and release notes carry artifact-specific status.