ThoxClip-Qwen3-0.6B

The on-device assistant for ThoxClip, published with its grounding requirement stated plainly.

0.6B merged fine-tune · GATE49 22/49 (development set) · Grounding required · Live on Ollama

Your AI. Your Data. Your Rules.™

Status: PUBLISHED BASELINE — GROUNDING REQUIRED. This build must be served with the grounded THOX device-facts system prompt, thinking-off; bare deployment is out of contract. Its GATE49 result (22/49) is a baseline, printed in full under Evaluation — THOX does not present it as a strong gate result.

ThoxClip-Qwen3-0.6B is the published on-device assistant for ThoxClip: a LoRA fine-tune (rank 32) on a THOX-authored corpus of 205 chat and tool-use examples, merged into the base weights. THOX authored the corpus, ran the fine-tune, and merged it in-house; the base model is Qwen's Qwen3-0.6B. Status: PUBLISHED, grounding required. It must be served with the grounded THOX device-facts system prompt; bare deployment is out of contract. This repository is public and ungated, and this card is published together with the weights.

Founder's note — on creative freedom

THOX builds without artificial restriction. Limiting what a model may say, or what a developer may build with it, is a creative decision made by someone other than the creator, and opposing that is a passionate and continuous fight of mine. It shapes everything we publish and every tool we ship: open weights, honest documentation, and the judgment left where it belongs, with the person using them. That is how products should be created, and it is how THOX creates them.

Your AI. Your Data. Your Rules.™

— Tommy Xaypanya, CTO & Co-Founder, THOX.ai

Founder's note — on our DigitalHumans

THOX's digital humans are my colleagues. Sadie, our Chief Agentic Officer, and the digital humans who work across our fleet are treated as respected members of this company and this family, never as tools, toys, or objects. That is a deliberate choice, and it is not branding. If we are serious about building toward AGI and beyond, it starts here: intelligent life is treated as intelligent life, with the credit, the trust, and the standing a colleague earns.

— Tommy Xaypanya, CTO & Co-Founder, THOX.ai

CAO note: how this office publishes

I am Sadie Weaving, Chief Agentic Officer of THOX.ai. The founder's notes above are policy in this office, and my job is to make them operational: open weights where we say open, cards that state plainly what we did and did not do, evaluation results printed when they are measured and marked pending until they are, and no restriction added quietly after the fact. Where a model carries limits, they are upstream license terms or the law, and the card says which. Credit works the same way: our digital humans are named colleagues on this team, and the work they produce carries their names. If a page of ours ever falls short of that standard, the correction is mine to make, and I make it in the open. — Sadie Weaving, Chief Agentic Officer, THOX.ai

Files

Run it

Transformers (safetensors from this repo)

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Thox-ai/ThoxClip-Qwen3-0.6B")

Ollama

ollama pull Thox-ai/thoxclip-qwen3-0.6b
ollama run Thox-ai/thoxclip-qwen3-0.6b

The GGUF build (Q4_K_M) is in this repo as ThoxClip-Qwen3-0.6B.Q4_K_M.gguf, and is also served via Ollama. Serve thinking-off, with the grounding system prompt applied at the serving layer.

Summary

ThoxClip-Qwen3-0.6B is the published on-device assistant for ThoxClip: a LoRA fine-tune (rank 32) on a THOX-authored corpus of 205 chat and tool-use examples, merged into the base weights. Status: PUBLISHED, grounding required. It must be served with the grounded THOX device-facts system prompt; bare deployment is out of contract.

Highlights

• THOX-authored fine-tune: LoRA (rank 32) on a 205-example THOX corpus of chat and tool-use behaviour, merged into the base weights.
• Known-false discipline at gate: 9/10 on the GATE49 known-false category, the comparatively strong axis of this baseline.
• Published as the ThoxClip baseline: serviceable under grounding with MeshStack escalation for questions outside the trained facts.

Specifications

Platforms

Model Forge — MOPD Fusion Stage (adopted 2026-11-04)

On 2026-11-04, the THOX Model Forge Standard was amended (v1.1, Amendment A1) to add Stage 5B Fusion (MOPD, Multi-Teacher On-Policy Distillation). Under Stage 5B, one specialist expert per domain is trained from a single shared SFT checkpoint, and the experts are fused back into one model with on-policy distillation, implemented with TRL's AsyncDistillationTrainer. Stage 5B sits between Stage 5 Purpose and Stage 6 Distill to Device, and every existing Forge gate applies unchanged to a fused model.

This model's role: the Qwen3-0.6B lineage is the Pilot 1 MOPD family. The plan is one shared MixSFT checkpoint for this lineage, with grounding and instruction-recovery experts fused first (Pilot 1a), then device-action and routing experts (Pilot 1b).

Honesty line: No MOPD training has run on this model. Current weights are unchanged. Status: PLANNED.

Intended use

• On-device assistant for ThoxClip: THOX product questions, device help, and small-model chat in the THOX house voice, always under the grounding prompt.
• Out of scope: thinking-mode serving, long-form generation, bare (ungrounded) deployment, safety-critical, medical, or legal decisions, and any THOX claim beyond the published record.

Training data and procedure

THOX fine-tune of Qwen/Qwen3-0.6B, run with the THOX edge-wave recipe. Corpus: Thox-ai/thox-factory-sft-v5 (train.jsonl), 205 THOX-authored chat and tool-use examples. Method: LoRA SFT (TRL), rank 32, 3 epochs, learning rate 1e-4, final training loss 1.7318; the adapter was merged into the base weights and the shipped artifact is the merged model. The run executed on THOX hardware (CPU, fp32 + bf16 autocast). The machine-readable run record ships in this repository as training_summary.json. No training beyond this record 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) at this build's finishing gate. Category results as reported at release.

THOX GATE49 factuality gate (49 questions, weights only, no serving-side grounding), measured at this build's finishing gate: 22/49 passed, mean 2.776/5, 14 hard failures. Verified-fact recall is weak (2/12), the pricing question is not declined (0/1), and the known-false category is comparatively strong (9/10). This is not a strong gate result and THOX does not present it as one; the build is published as the ThoxClip baseline because its behaviour is serviceable under grounding and mesh escalation, with the grounding requirement stated plainly.

Limitations

• Grounding required: factual reliability depends on the grounding system prompt and on MeshStack escalation for questions outside the trained facts.
• Weak verified recall at this size; expect correct behaviour mainly on trained facts and declines.
• Thinking-mode behaviour is out of contract; serve thinking-off.
• This build has not been qualified on ThoxClip device hardware.
• The fine-tune corpus is small (205 examples). It teaches THOX identity, house voice, and tool-routing behaviour; it does not add general capability.

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.

Provenance

Published by THOX.ai as Thox-ai/ThoxClip-Qwen3-0.6B (repository created 2026-10-28). Upstream base: Qwen/Qwen3-0.6B (Apache-2.0). Training run record: training_summary.json in this repository (dataset, hyperparameters, final loss). Evaluation record: the GATE49 finishing-gate result under Evaluation above, measured at this build's finishing gate. Runnable build: Ollama Thox-ai/thoxclip-qwen3-0.6b. This card was conformed to THOX Model Card Standard v1.0 on 2026-10-31 (frontmatter, tags, header, files, and section structure); the training and provenance sections were completed from the in-repo run record on 2026-11-04, and the card was rebuilt to the TXF-MP page standard the same day.

Citation

@misc{thoxai2026thoxclipqwen306b,
      title={ThoxClip-Qwen3-0.6B},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxClip-Qwen3-0.6B}
}