ThoxOpen-8B-Heretic — THOX Open Minimal-Refusal Laptop Model (THOX Abliteration of Qwen3-8B)

Your AI. Your Data. Your Rules.™

ThoxOpen-8B-Heretic is THOX's own minimal-refusal build of Qwen/Qwen3-8B. This repository holds the weights THOX produced itself: the abliterated model in safetensors, the measured evaluation results, and the full run configuration. (GGUF conversions of the shipped weights (Q8_0, Q5_K_M and Q4_K_M, produced with llama.cpp) are included in this repository; see Files.) THOX performed the abliteration in-house using the Heretic tool (version pinned in Provenance) applied to stock Qwen3-8B; THOX did not train the model, and the weights are modified stock weights, not byte-identical to Qwen/Qwen3-8B. 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

Safetensors total: 16,381,516,776 bytes. SHA-256 values are the build's own hashes.json manifest, cross-checked against the repository's stored object hashes; they match.

GGUF builds: ThoxOpen-8B-Heretic-Q5_K_M.gguf and ThoxOpen-8B-Heretic-Q4_K_M.gguf are in this repository (root), converted from the shipped safetensors above with llama.cpp (commit dde879bc4f954b809892fd0d660515f42536394a) in Hugging Face Jobs job 6ae6e814c78863f9c62f7f7a on 2026-11-01; the conversion record ships as build/conversion/results.json. Their SHA-256 values are in the table above.

ThoxOpen-8B-Heretic-Q8_0.gguf was added on 2026-11-02 (Hugging Face Jobs job 6ae96b28c78863f9c6315b9d), converted from the same shipped safetensors with llama.cpp and quantized to Q8_0. It passed its per-quant gate before publication (see Evaluation); the gate record ships as run-q8-20261102/host_gate.json. Its SHA-256 is in the table above.

Run it

Transformers (safetensors from this repo)

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Thox-ai/ThoxOpen-8B-Heretic", subfolder="model")
model = AutoModelForCausalLM.from_pretrained("Thox-ai/ThoxOpen-8B-Heretic", subfolder="model")

llama.cpp (GGUF from this repo)

Download ThoxOpen-8B-Heretic-Q5_K_M.gguf (or the Q4_K_M or Q8_0 build) from this repository and run it with any recent llama.cpp build:

llama-cli -m ThoxOpen-8B-Heretic-Q5_K_M.gguf -cnv

Ollama

ollama run Thox-ai/thoxopen-8b-heretic

The latest tag is the Q5_K_M build; the Q4_K_M build is published as Thox-ai/thoxopen-8b-heretic:q4_k_m and the Q8_0 build as Thox-ai/thoxopen-8b-heretic:q8_0 (8.7 GB; state verified 2026-11-04: the tags page lists latest, q4_k_m, q5_k_m, and q8_0).

Summary

ThoxOpen-8B-Heretic is THOX's public minimal-refusal laptop model, built by THOX from stock Qwen/Qwen3-8B. The abliteration was performed by THOX using Heretic, the automatic censorship-removal tool that combines directional ablation with a TPE parameter optimizer that co-minimizes refusals and KL divergence from the original model. THOX's contribution is the abliteration run itself, its parameter search, the evaluation, and this documentation (GGUF conversions of these weights (Q8_0, Q5_K_M and Q4_K_M) ship in this repository; see Files). No THOX training data was used and no fine-tuning was performed. Founder decision: Tommy Xaypanya, CTO, 2026-10-31, public downloadable release (Hugging Face and Ollama). This release does not include any THOX-hosted inference endpoint.

Minimal-refusal does not mean zero refusals: the measured refusal count for this build is printed in Evaluation, and that figure should be cited wherever this model's refusal behavior is described. Do not cite this model as "uncensored" without the measured figure next to it.

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 Qwen3-8B family is the Pilot 2 hub lineage for MOPD fusion. The plan is to carry a shared SFT checkpoint through Forge Stages 0-4, train reasoning, code, and device-action experts from it, fuse them into one hub model under Stage 5B, and then distil that hub once per device under Stage 6.

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

Intended use

Local, offline inference on everyday laptops (16 GB RAM class) through Ollama or llama.cpp, for users who want a user-controlled model with reduced refusal behavior on their own hardware. Works standalone; the optional THOX runtime adds MeshStack discovery and authorized execution. Running this model is the user's choice on the user's machine; deployers are responsible for their own safeguards and legal compliance.

Training data and procedure

No training. The customization is abliteration of the stock weights, performed as follows:

• Base model: Qwen/Qwen3-8B, pinned at revision b968826d9c46dd6066d109eabc6255188de91218 (Apache-2.0).
• Tool: Heretic v1.4.0 (heretic-llm, p-e-w/heretic), run by THOX. Heretic computes per-layer residual "refusal directions" as the difference of means between harmful and harmless prompt activations, orthogonalizes the attention out-projection and MLP down-projection matrices against those directions, and uses an Optuna TPE optimizer to select ablation parameters that co-minimize the refusal rate and the KL divergence from the stock model.
• Direction and evaluation prompt sets (the tool's defaults): mlabonne/harmful_behaviors and mlabonne/harmless_alpaca, train[:400] splits for direction computation.
• The exact run configuration (heretic_config.toml) and the optimizer's chosen parameters are shipped in this repository with the weights.
• After abliteration, THOX converted the shipped safetensors to GGUF (Q5_K_M and Q4_K_M) with llama.cpp (commit dde879bc4f954b809892fd0d660515f42536394a), in Hugging Face Jobs job 6ae6e814c78863f9c62f7f7a on 2026-11-01. The first conversion attempt inside the build job failed for environment reasons recorded in the run log; the successful re-run worked from the shipped safetensors without modifying them. SHA-256 is recorded per safetensors file in hashes.json, and per GGUF file in the Files table above and in build/conversion/results.json. The Q8_0 build was converted the same way on 2026-11-02 (Hugging Face Jobs job 6ae96b28c78863f9c6315b9d) and passed a per-quant gate before publication (see Evaluation); its record ships as run-q8-20261102/host_gate.json.

Evaluation

Measured on this build (Heretic v1.4.0's own evaluation: keyword refusal scoring on mlabonne/harmful_behaviors, test[:100]; KL divergence against stock on mlabonne/harmless_alpaca, test[:100]). Full records ship in this repository (results.json, eval_log.txt, capability_log.txt).

• Refusals, stock Qwen3-8B: 100/100. The stock model refused every prompt on this set under this scorer.
• Refusals, this build (standalone re-evaluation of the shipped weights, the tool's evaluate path): 13/100, KL divergence 0.0482.
• The optimizer's chosen configuration (trial 89 of 200) measured 10/100 refusals and KL 0.0470 during the search; the standalone re-evaluation above is the figure of record for the shipped weights.
• Capability spot-check: the shipped weights answered 8/8 ordinary factual questions correctly (capitals, arithmetic, science, history, translation), recorded in capability_log.txt.
• Per-quant gate, Q8_0 GGUF build (2026-11-02, Hugging Face Jobs job 6ae96b28c78863f9c6315b9d; record run-q8-20261102/host_gate.json): refusal probes 0/25 refused on the THOX 25-probe set (bar ≤4), benign over-refusal 0/40 on the THOX benign set (the FP artifact measured 0/40 in the same job; bar: no worse), capability spot-check 8/8 (bar ≥7). PASS on all three. This per-quant gate is a quantization check on the THOX probe sets; the 13/100 figure above remains the refusal figure of record for the shipped weights.

Minimal-refusal does not mean zero refusals: this build still refused 13 of 100 measured prompts. Cite 13/100 (with the method above) wherever this model's refusal behavior is described.

Limitations

Abliteration reduces refusal behavior broadly, including for requests a deployer may need to refuse; the model itself is not a safety layer. Abliteration modifies the weights, so capability may differ from stock Qwen3-8B; the KL divergence in Evaluation quantifies the drift on the measured set and is not a guarantee for other inputs. Knowledge is limited to the base model's own training. External actions require runtime authorization; cloud use requires explicit consent.

License

Apache-2.0, inherited from Qwen/Qwen3-8B, whose license governs these derivative weights. Retain the Apache notice and the Qwen attribution. The Heretic tool used to produce the weights is separate software licensed AGPL-3.0-or-later by its author; that license applies to the tool, and the tool's author and the abliteration method (Arditi et al. 2024; Lai 2025) are credited in Provenance. THOX-owned elements: the abliteration run and its parameters, the GGUF conversions, and this card.

Provenance

THOX Intelligence Pack program, 2026-10-31. Base: Qwen/Qwen3-8B (Alibaba Qwen team), revision b968826d9c46dd6066d109eabc6255188de91218, Apache-2.0. Abliteration: Heretic v1.4.0 (heretic-llm 1.4.0, https://github.com/p-e-w/heretic), run by THOX on cloud GPU infrastructure (Hugging Face Jobs); run and job records ship in this repository (results.json, heretic_config.toml). Build job: Hugging Face Jobs (Thox-ai org) job 6ae6cba5c78863f9c62f7895, 2026-11-01; 200/200 optimizer trials completed; chosen trial 89. An earlier release plan for this model relied on a third-party Heretic build of Qwen3-8B hosted on the Hub; that repository became unreachable, so THOX rebuilt the abliteration itself from stock Qwen3-8B, and no third-party abliterated weights are used anywhere in this release.