ThoxNano-360M
ThoxClip Edge Assistant — an on-device assistant model for the ThoxClip tier: a LoRA fine-tune (rank 32) on a THOX-authored corpus of 982 chat, tool-use, and factuality examples, merged into the base weights
GATE49 23/49 (development set) · Safetensors weights · GGUF: Q4_K_M · Live on Ollama
Your AI. Your Data. Your Rules.™
Status: PUBLISHED — GROUNDING REQUIRED. Deployed surfaces supply facts through the serving layer; the bare weights are not a reliable source of THOX facts. The full gate result is printed under Evaluation.
Summary
ThoxNano-360M is an on-device assistant model for the ThoxClip tier: a LoRA fine-tune (rank 32) on a THOX-authored corpus of 982 chat, tool-use, and factuality examples, merged into the base weights. Status: PUBLISHED, grounding required. Deployed surfaces supply facts through the serving layer; the bare weights are not a reliable source of THOX facts.
Highlights
- THOX fine-tune: a LoRA fine-tune (rank 32) on a THOX-authored corpus of 982 chat, tool-use, and factuality examples, merged into the base weights.
- GATE49 23/49 on the THOX factuality gate (development set, abstention-weighted), mean 2.959/5, 11 hard failure(s) — printed in full under Evaluation.
- This repo carries the merged weights (safetensors), a Q4_K_M GGUF, and the chat template (
chat_template.jinja, shipped in this repo).
Specifications
| Base model | HuggingFaceTB/SmolLM-360M |
| Parameters | 360M |
| Formats | Safetensors weights · GGUF: Q4_K_M · Ollama Modelfile |
| Training | LoRA fine-tune (rank 32) on a THOX-authored corpus of 982 chat, tool-use, and factuality examples, merged into the base weights |
| License | Apache-2.0. The THOX fine-tune and additions in this repo are covered under Apache-2.0; see the license text for terms. |
| This repository contains | Full weights and tokenizer in this repo |
Files
| File | Size | SHA-256 |
|---|---|---|
model.safetensors |
723,674,912 B | eda05372b511be9d809d7bc65dc93c847c6f1d28f598b7fc9abcfeaf32c1a758 |
ThoxNano-360M.Q4_K_M.gguf |
270,590,336 B | b1c6dcc9c91a8664c420edcc9dbd9f47967f37b5abe193a2a4f5a37ecd904876 |
tokenizer.json / .gitattributes / tokenizer_config.json / Modelfile / config.json / training_summary.json / chat_template.jinja / generation_config.json |
— | Supporting files (see repo) |
Run it
This repo carries the merged weights (safetensors), a Q4_K_M GGUF, and the chat template (chat_template.jinja, shipped in this repo).
ollama pull Thox-ai/thoxnano-360m
ollama run Thox-ai/thoxnano-360m
hf download Thox-ai/ThoxNano-360M ThoxNano-360M.Q4_K_M.gguf --local-dir ./ThoxNano-360M
llama-cli -m ./ThoxNano-360M/ThoxNano-360M.Q4_K_M.gguf -p "Hello" -n 128
Platforms
| Platform / device | Run path |
|---|---|
| Ollama (laptop / desktop) | ollama run Thox-ai/thoxnano-360m |
| Hugging Face | Download the builds listed under Files from this repo (see Run it) |
Intended use
- On-device assistant and tool routing for ThoxClip: short conversational replies, THOX house voice, and routing intents to device and mesh functions, fully offline, with facts supplied by the serving layer.
- Out of scope: safety-critical, medical, or legal decisions without a human in the loop, and any use as a standalone source of THOX facts.
Training data and procedure
THOX fine-tune of HuggingFaceTB/SmolLM-360M. The corpus and procedure are described in this card as published at release (see Summary and Evaluation); no training beyond what this card describes is claimed for this repository.
Evaluation
Measured with the THOX GATE49 factuality gate (49 questions, weights only, no serving-side grounding; GATE49 is a development set, abstention-weighted). Full narrative under Evaluation below.
| Benchmark | Result |
|---|---|
| Total | 23/49 — mean 2.959/5 |
| Hard failures | 11 |
THOX GATE49 factuality gate (49 questions, weights only, no serving-side grounding), most recent evaluation of the shipped build: 23/49 passed, mean 2.959/5, 11 hard failures. This is a weak gate result. An earlier evaluation run could not complete because the model fell into a repetition loop on a gate question; that tendency is recorded under Limitations. The build is published for identity, voice, and routing behaviour under grounding, not for factual recall from the weights.
Limitations
- The model can fall into repetition loops on some prompts; keep output limits modest and stop conditions on.
- It can echo fragments of its system text instead of answering conversationally.
- Factual answers from the bare weights are unreliable; deploy with serving-side grounding for facts.
- Quantization trades accuracy for size; behaviour and bias reflect the training data, which is mostly English, and quality in other languages is lower.
- This build has not been qualified on ThoxClip device hardware.
License
Apache-2.0. The THOX fine-tune and additions in this repo are covered under Apache-2.0; see the license text for terms.
THOX.ai — Your AI. Your Data. Your Rules.™
Provenance
Published by THOX.ai as Thox-ai/ThoxNano-360M. Upstream base: HuggingFaceTB/SmolLM-360M. This card was conformed to THOX Model Card Standard v1.0 on 2026-10-31 (frontmatter, tags, header, files, and section structure); the technical content is as published at release.
Citation
@misc{thoxai2026thoxnano360m,
title={ThoxNano-360M},
author={THOX.ai},
year={2026},
url={https://huggingface.co/Thox-ai/ThoxNano-360M}
}