ThoxHeretic-9B

⚠️ Safety notice. This repo redistributes a community uncensored / abliterated model merge. It is not THOX-trained and not safety-tuned; refusal behaviour is deliberately reduced by its upstream authors. It is not part of the ThoxOS or MeshStack catalog and is not promoted for THOX devices. It is public and ungated by deliberate founder exception (the Heretic family posture, set for ThoxHeretic-27B on 2026-10-31 and carried here). Do not deploy it to end users without your own safety layer.

A 9B creative-writing and reasoning merge from the community Qwen3.5-9B “Heretic” line, distributed by THOX.ai as a Q4_K_M GGUF on Ollama with a paired vision projector, a THOX serving configuration, and a THOX system prompt. The weights are third-party community work; THOX did not train, merge, or abliterate them (see “What THOX Did and Did Not Do”).

Model ID: Thox-ai/ThoxHeretic-9B Base model: Qwen3.5-9B by Qwen (ultimate base of a 6-entry community merge lineage, listed under Provenance) Runnable build: Ollama - ollama run Thox-ai/thoxheretic:9b License: Apache-2.0 (see License Terms)

Your AI. Your Data. Your Rules.™

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

What THOX Did and Did Not Do

This is a THOX-configured distribution of community weights, and the THOX work lives in the serving layer, never in the weights.

What THOX did: selected and staged this GGUF build, authored the Ollama Modelfile (a THOX ChatML template override, sampling parameters, and stop tokens), authored the THOX system prompt the Ollama serving carries (an identity/framing prompt; it does not change the weights’ behaviour class), paired the Qwen3.5 vision projector (mmproj) into the Ollama distribution by manifest, and published and documented the distribution — publicly and ungated, under the Heretic family’s deliberate founder exception.

What THOX did not do: THOX did not train, fine-tune, merge, or abliterate these weights, and did not safety-tune them. The weights are a community mergekit merge of Qwen3.5-9B fine-tunes by Nightmedia and DavidAU, several of them abliterated “Heretic” tunes (full lineage under Provenance). THOX has not run a benchmark or evaluation suite on this artifact; see Evaluation.

Specifications

Intended Use

• Creative writing, roleplay, fiction, and brainstorming on the operator’s own hardware.
• Uncensored reasoning and research/evaluation of abliterated community models, behind the operator’s own guardrails.

Out of scope: end-user deployment without a safety layer; medical, legal, financial, or safety-critical decisions without qualified human review; unattended control of physical devices or actuators without a confirmation step; any use prohibited by the upstream licence.

How to Use

ollama pull Thox-ai/thoxheretic:9b
ollama run Thox-ai/thoxheretic:9b

This Hugging Face repository is the model’s public model card page. The runnable build is served from Ollama at https://ollama.com/Thox-ai/thoxheretic — the same facts and framing appear on that page.

Training Data and Procedure

The weights are a community merge (mergekit) of the fine-tunes listed under Provenance, produced by their respective authors from Qwen/Qwen3.5-9B. Several components are “Heretic” abliterated tunes — models whose refusal directions were ablated upstream — which is why this merge’s refusal behaviour is reduced. Their training data and procedures are those of the upstream authors and the base model’s provider; see the upstream repos and the base model card for that record.

Datasets added by THOX for this release: none. THOX fine-tuning or ablation runs on these weights: none.

Evaluation

Independent THOX evaluation of this model: pending. No benchmark, gate, or vision evaluation has been run by THOX on this artifact, and this card reproduces no upstream performance figures. This section will be updated with method and results when a THOX-run evaluation lands.

Release-process checks that did run: the publish pipeline’s generation smoke (the model must generate through its served template before any push) and the distribution build checks recorded in the release report. Those are serving checks, not quality evaluations.

Limitations

• Uncensored / abliterated merge: the model will comply with requests a safety-tuned model would refuse. The operator is responsible for the deployment context and any guardrails.
• Behaviour and bias reflect the upstream authors’ training data and the merge; output quality varies by topic and language (GGUF metadata lists English and Chinese).
• Vision is provided by grafting the upstream Qwen3.5 projector onto a merged language model whose components were not jointly vision-tuned by THOX; image understanding quality is unevaluated by THOX.
• Served context defaults to 32,768 tokens regardless of the larger GGUF-declared window; long-context behaviour is unevaluated.
• All processing is local by design; the model has no access to online information or external services unless the operator wires them in.
• Outputs can be wrong. Professional, medical, legal, financial, and security-relevant outputs require qualified human review.

License Terms

The weights in this distribution are declared Apache License 2.0: the GGUF metadata embedded in the artifact records general.license: apache-2.0, and the ultimate base model, Qwen/Qwen3.5-9B, is published under Apache-2.0. THOX claims no license over the weights beyond the upstream terms. The intermediate community repositories in the merge lineage carry their own notices; verify them before any redistribution. THOX’s own material in this distribution (the Modelfile configuration, system prompt, and this documentation) is THOX.ai LLC material and does not extend any additional restriction to the weights.

Provenance

• GGUF embedded identity: Qwen3.5 9B DS9 USS Defiant NM (finetune tag DS9-USS-Defiant-NM).
• Merge lineage recorded in the GGUF, in order: Qwen3.5 9B TNG PKD Qwopus Coder (Nightmedia); Qwen3.5 9B TNG PKD Qwopus Coder Fable Polaris Writer V4 ANDR2B (Nightmedia); Qwen3.5 9B TNG PKD Qwopus Writer Bradbury F451 Orwell 1984 (Nightmedia); Qwen3.5 9B Pro Writer 1984 Orwell Uncensored Heretic (DavidAU); Qwen3.5 9B The Bradbury F451 Pro Writer Uncensored Heretic (DavidAU); Qwen3.5 9B Polar Rev1 Uncensored Heretic (DavidAU). Ultimate base: Qwen/Qwen3.5-9B.
• Publicly released 2026-10-31 by order of Tommy Xaypanya, CTO and Co-Founder of THOX.ai, as part of the staged-models release.
• Runnable build: Ollama, ollama run Thox-ai/thoxheretic:9b and https://ollama.com/Thox-ai/thoxheretic
• This Hugging Face repository was created 2026-10-31 as the model’s public model card page. It contains this card; the runnable model build is served from Ollama.
• Sibling artifact, distinct from this one: ThoxHeretic-27B (https://huggingface.co/Thox-ai/ThoxHeretic-27B), a larger DavidAU Qwen3.6-27B Heretic redistribution under the same family posture. Facts and figures are not interchangeable between the two.

Citation

@software{thoxai_thoxheretic_9b_2026,
  title = {ThoxHeretic-9B},
  author = {Thox.ai LLC},
  year = {2026},
  url = {https://huggingface.co/Thox-ai/ThoxHeretic-9B},
  note = {Runnable build: https://ollama.com/Thox-ai/thoxheretic. Weights: community Qwen3.5-9B Heretic merge lineage; see Provenance.}
}

Thox.ai LLC — Your AI. Your Data. Your Rules.™