ThoxNeedle-Micro

On-Device Function Calling for THOX Hardware — a THOX-trained function-calling model for THOX hardware, packaged in the .cact format the device runtime loads

Function-calling for THOX device tools · .cact device format (13.7 MB) · Empty-call discipline for unservable requests

Your AI. Your Data. Your Rules.™

Status: PUBLISHED. A THOX-trained function-calling model for THOX hardware, shipped in the .cact device format (not GGUF). Held-out evaluation is printed in full under Evaluation. See Files for exactly what this repo contains.

Summary

ThoxNeedle-Micro is a THOX-trained function-calling model for THOX hardware, packaged in the .cact format the device runtime loads. Given a set of THOX device tool schemas and a user request, it emits the tool call that satisfies the request, or an empty call list when no available tool can serve it. That second behaviour is a first-class objective: on a device, a model that invents a plausible-looking call for an unservable request fires real actuators.

Highlights

• ThoxNeedle-Micro is a THOX-trained function-calling model for THOX hardware, packaged in the .cact format the device runtime loads.
• This repo ships .cact weights, not GGUF; run them with the Cactus runtime on a ThoxOS phone or host.

Specifications

Files

Run it

This repo ships .cact weights, not GGUF; run them with the Cactus runtime on a ThoxOS phone or host. There is no llama.cpp / Ollama path for .cact files.

hf download Thox-ai/ThoxNeedle-Micro --include "*.cact" --local-dir ./ThoxNeedle-Micro

Platforms

Training data and procedure

LoRA fine-tune (rank 16) on a THOX-authored corpus of 1,400 device tool-calling examples, split group-aware (1,145 train / 255 held out) so no example appears on both sides. The fine-tune's changes survive the deployment quantization: 58.38% of weight elements changed versus the passthrough build, verified by dequantizing and diffing both artifacts.

Evaluation

Held-out set: 255 distinct examples, greedy decoding. Exact match and refusal accuracy are reported separately because averaging them would hide the failure mode that matters on a device.

Read these deltas with the sample size in mind: one example is 0.39 percentage points, so single-example movements are not trends this set can resolve. No THOX GATE49 result is published for this model.

Intended use

• Local, on-device function calling for THOX hardware against THOX device tool schemas.
• Out of scope: medical, legal, financial, safety-critical, or other high-stakes decisions without human review; unattended control of physical devices without a confirmation step.

Limitations

• The training corpus is small (1,400 examples), which caps what any fine-tune of it can demonstrate; expanding it is the prerequisite for a meaningful improvement.
• Tool schemas outside the trained THOX device set are out of scope; the model should refuse them, and refusal accuracy is 89.86%, not 100%.
• Behaviour and bias reflect the training data, which is mostly English; quality in other languages is lower.

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/ThoxNeedle-Micro. THOX from-scratch research line; no upstream base model. 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{thoxai2026thoxneedlemicro,
      title={ThoxNeedle-Micro},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxNeedle-Micro}
}