THOX.ai LLC

company

https://www.THOX.ai

AI & ML interests

At THOX.ai, we build local-first, privacy-first AI that runs where users need it most: on edge devices, workstations, portable hardware, and embedded systems. Our research and engineering focus on making advanced AI practical without requiring cloud infrastructure or sacrificing user ownership of data. Areas of Interest * Edge AI and on-device inference * Local-first LLM deployment * Small Language Models (SLMs) * Efficient transformer architectures * Quantization and model optimization * GGUF, LiteRT, and embedded AI runtimes * AI acceleration on consumer hardware * Mobile and embedded AI systems * Privacy-preserving AI * Offline AI assistants * Agentic AI systems * Multi-agent orchestration * AI operating systems * Retrieval-Augmented Generation (RAG) * Semantic search and knowledge graphs * Memory architectures for AI agents * AI developer tools * AI infrastructure * Open-source AI * Human-AI collaboration * AI for healthcare * AI for education * AI for accessibility * AI for legal and enterprise workflows * Robotics and autonomous systems * Digital humans * Computer vision * Speech and multimodal AI * Federated and distributed AI * Edge-to-edge AI networking * Secure AI deployment * Model benchmarking and evaluation * AI hardware integration * Quantum-inspired optimization * Responsible AI engineering Technologies We Explore * Gemma * LiteRT * llama.cpp * GGUF * ONNX Runtime * TensorFlow Lite * PyTorch * Hugging Face Transformers * Rust * C++ * Python * WebGPU * CUDA * Vulkan * Jetson * Raspberry Pi * Embedded Linux Our Mission THOX.ai develops open, modular AI technologies that help developers, researchers, businesses, and makers deploy powerful AI locally. We believe users should have meaningful control over their models, data, and computing resources while benefiting from modern AI capabilities across desktop, mobile, embedded, and edge environments.

Recent Activity

tommytracx published a model 4 days ago

Thox-ai/fable5-agentic-qwen2.5-14b-lora

tommytracx updated a model 5 days ago

Thox-ai/fable5-agentic-qwen2.5-14b-lora

tommytracx updated a Space 5 days ago

Thox-ai/ThoxMythos-9B-Space

View all activity

Organization Card

Community About org cards

THOX.ai — On-device intelligence

THOX ships small, focused models tuned for the THOX device family (ThoxKey, ThoxClip, ThoxMini, ThoxAir, ThoxNova). Weights, fine-tunes, and quantizations live here on Hugging Face; GGUF and Ollama tags mirror on ollama.com/Thox-ai. The stack is built around thoxllm-factory v0.1.6 and the Phase C Gemma-4 12B line that shipped 2026-07-02. Kickstarter launch 2026-08-12.

Profile: huggingface.co/Thox-ai · Website: thox.ai · GitHub: github.com/ttracx · License defaults to Apache-2.0 for THOX-owned artifacts; derivatives inherit their base license.

What we ship

THOX.ai owns the stack end to end. Three layers:

Foundation models trained from scratch on THOX-defined architectures — not fine-tunes of someone else's base (ThoxLLM-327M line).
Adapters and instruction tunes over community bases (Llama, Mistral, Qwen, Gemma) via thoxllm-factory, sized for specific devices.
Edge devices the models run on: ThoxKey, ThoxClip, ThoxMini, ThoxAir, ThoxNova. See github.com/ttracx/thox-portable.

Model families

Micro line — edge / MCU

Model	Base	Notes
ThoxLLM-327M	from scratch	THOX single-projection SwiGLU, GQA-16:8, RoPE, 20L / 1024. Research seed, WikiText-103, 50k steps.
ThoxLLM-327M-v2	from scratch	Path A production variant, standard 2-projection SwiGLU (GGUF-clean), 100k steps, `thoxllm-v2` corpus.
thox-micro-125m	from scratch	123.6M decoder-only, val 2.61. Coming soon to HF.

Compact line — ThoxMini / ThoxAir

Model	Base	Notes
ThoxMini-3B	`meta-llama/Llama-3.2-3B-Instruct`	THOX-SFT via factory `run3`, 240 steps.
ThoxForge-7B	`mistralai/Mistral-7B-Instruct-v0.3`	THOX-SFT.
ThoxGlobal-7B	`Qwen/Qwen2.5-7B-Instruct`	THOX-SFT, multilingual.
thox-gem-e4b	`unsloth/gemma-4-e4b-it-unsloth-bnb-4bit`	Gemma-4 E4B QLoRA adapter, factory `rig-4060ti-e4b-qlora`.

Wave line — ThoxNova

Model	Base	Notes
ThoxWave-8B-Unleashed	`cognitivecomputations/dolphin-2.9-llama3-8b`	Dolphin-Llama-3 SFT LoRA.
ThoxNova-12B-Unleashed	`MarinaraSpaghetti/NemoMix-Unleashed-12B`	Mistral NemoMix 12B SFT LoRA.
ThoxNova-12B-Agent	Gemma-4 12B (Phase C v0.2.1)	Agent SFT. Coming soon to HF.
ThoxNova-12B-Core	Gemma-4 12B (Phase C v0.2.1)	Core SFT. Coming soon to HF.

Gemma-4 base — Phase C, 2026-07-02

Model	Base	Notes
thoxgemma4-12b-lora	`unsloth/gemma-4-12b-it`	QLoRA adapter (~131 MB), factory `rig-4060ti-12b-qlora`.
thox-gemma4-12b-merged	Gemma-4 12B	Merged bf16 (~23.8 GB). Coming soon.
thox-gemma4-12b-gguf	Gemma-4 12B	Q4_K_M (~~6.87 GB) + f16 (~~23.83 GB). Coming soon; mirrors to Ollama.

Where to run

Ollama — ollama pull thox-ai/<name>:<tag> from ollama.com/Thox-ai. Recommended for on-device (Nova, Mini, Air).
llama.cpp / LM Studio — pull GGUF directly from a model repo's Files tab once GGUF variants ship.
Native transformers — AutoModelForCausalLM.from_pretrained("Thox-ai/<name>") on transformers >= 5.6.0. Private repos require a Hugging Face token with read access.
Adapters — PEFT / LoRA adapters need the base model at inference time. See the individual model card for the exact base_model: and merge steps.

Devices

Models are tuned against the THOX device fleet: ThoxKey (USB fob), ThoxClip, ThoxMini, ThoxMini Air, ThoxNova. Device specs and boot images live at github.com/ttracx/thox-portable.

License

Per model. The base-model heritage determines the terms:

ThoxLLM-* — Apache-2.0 (THOX-owned end-to-end).
Llama-derived (ThoxMini-3B, ThoxWave-8B-Unleashed) — inherits the Meta Llama Community License (3.2 and 3 respectively).
Gemma-derived (thox-gem-e4b, thoxgemma4-12b-lora, and the coming ThoxNova-12B-Agent/Core, thox-gemma4-12b-merged/gguf) — inherits the Gemma Terms of Use.
Mistral-derived (ThoxForge-7B, ThoxNova-12B-Unleashed) — Apache-2.0.
Qwen-derived (ThoxGlobal-7B) — inherits Tongyi Qianwen License.

Always check the individual model card for the authoritative terms.