THOX.ai
Edge and on-device language models. We build small models that run on hardware people actually own — laptops, USB sticks, single-board computers, and microcontrollers.
14 models are published here. Everything listed below is real and downloadable; nothing on this page links to a repository you cannot open.
Open — no access request
| model | what it is | size |
|---|---|---|
| ThoxMicro-1bit-9M | BitNet b1.58 ternary LM, trained from scratch. Writes children's stories; not an assistant. Research artifact for microcontroller-class targets. | 6.3 MB |
| ThoxMicro-1bit-16M | The same model at 16 layers instead of 8. 10.7% lower perplexity; needs more device headroom. Neither supersedes the other. | 8.0 MB |
| thox-micro-125m · GGUF | 123.6M base LM trained from scratch. Completion only. | 81–247 MB |
| ThoxSpeech-ASR-BitNet | BitNet speech recognition, redistributed byte-identical with published SHA-256. | 0.7–1.0 GB |
Access-gated
Adapters and derived models. Request access on the model page.
| model | base | type |
|---|---|---|
| ThoxMini-3B | Llama-3.2-3B-Instruct | LoRA + GGUF |
| ThoxForge-7B | Mistral-7B-Instruct-v0.3 | LoRA + GGUF |
| ThoxGlobal-7B | Qwen2.5-7B-Instruct | LoRA + GGUF |
| ThoxWave-8B | dolphin-2.9-llama3-8b | LoRA + GGUF |
| ThoxNova-12B-Core · Agent | Gemma-4-12B-it | LoRA + GGUF |
| fable5-agentic-qwen2.5-14b-lora | Qwen2.5-14B-Instruct | QLoRA adapter |
On the adapter bases: ThoxMini, ThoxForge, ThoxGlobal and both ThoxNova adapters were trained against
unsloth/*re-uploads of the listed models, not the original repositories. Each card'sbase_modelnow states the re-upload it was actually trained on, with the upstream lineage recorded in the body. ThoxWave-8B trains directly ondphn/dolphin-2.9-llama3-8b.
Documentation-only
These contain no weights, deliberately, and say so on their own pages:
- ThoxForge-7B-GGUF — the quantized build did not pass our factuality gate; publishing it anyway to make the repo look complete would route around that control.
- ThoxWebby-Gemma-4-E2B — the browser app pulls upstream weights; we do not re-host them.
Where our models run
- Ollama —
ollama.com/thox-ai - llama.cpp — every GGUF above
- ThoxKey / ThoxKey Pro — USB sticks that carry a model set and a local runtime
- ThoxMicro / ThoxOS devices — ESP32-S3 and Raspberry Pi class hardware
How we publish
- Measured, not estimated. Numbers on our cards come from runs we did. Where a figure is inherited from an upstream benchmark, the card says so.
- Negative results ship too. ThoxMicro-1bit-9M's card leads with the fact that ternary quantization costs it 21.4% perplexity, and that the penalty grew rather than shrank with more training data.
- Empty is better than misleading. A repo with no weights and an honest explanation beats one with weights that failed their gate.
- Licence honesty. Where downstream terms are genuinely unsettled we say
license: otherand explain, rather than picking a permissive licence because it is convenient.