ThoxEdge-0.8B

The ThoxEdge flagship: a merged 0.8B build that answers from canonical THOX facts and declines what THOX has not published.

GATE49 46/49 (development set) · Merged bf16 weights · Q4_K_M + Q8_0 GGUF · Live on Ollama

Your AI. Your Data. Your Rules.™

Status: PROMOTED flagship — passed the publish bar: GATE49 46/49 (development set), zero known-false hard failures. Measured at the Wave C finishing gate; see Evaluation.

ThoxEdge-0.8B is the merged-weights build of the Wave C ThoxEdge flagship: the LoRA adapter in Thox-ai/ThoxEdge-0.8B-WaveC merged into its base, Qwen/Qwen3.5-0.8B (Apache-2.0), in bf16. THOX trained the adapter on its own SFT data and merged it in-house; the base model is Qwen's. This repo holds the servable form of the adapter — merged weights, GGUF conversions, and the Ollama Modelfile. The Wave B GGUF for this tier was never published (its gate read was a serving artifact), so this repo is also the tier's first published merged build on the 982-row corpus. 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

Ollama

ollama run Thox-ai/thoxedge-0.8b

Hugging Face (GGUF download)

hf download Thox-ai/ThoxEdge-0.8B ThoxEdge-0.8B.Q4_K_M.gguf --local-dir ./ThoxEdge-0.8B

Run in Llama (macOS / Windows)

Run ThoxEdge-0.8B locally with Llama (by the llama.cpp team and Hugging Face) — or with the llama.cpp command line, which is the same engine:

llama serve -hf Thox-ai/ThoxEdge-0.8B:Q4_K_M --port 9931

Then chat in the app, or call the OpenAI-compatible API:

curl http://localhost:9931/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Thox-ai/ThoxEdge-0.8B:Q4_K_M",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Fallback (exact file, no resolver involved):

hf download Thox-ai/ThoxEdge-0.8B ThoxEdge-0.8B.Q4_K_M.gguf --local-dir ./ThoxEdge-0.8B
llama serve -m ./ThoxEdge-0.8B/ThoxEdge-0.8B.Q4_K_M.gguf --port 9931

RAM totals are calculated estimates (GGUF file size plus runtime buffers and KV cache), not runtime measurements.

Optional models.user.ini override (macOS; the app regenerates its own config on every launch, user overrides win):

[Thox-ai/ThoxEdge-0.8B:Q4_K_M]
temp = 0
ctx-size = 4096

Honest notes:

• Thinking-off does not transfer. The gated Ollama build serves this model thinking-off via the Modelfile's empty-think prefill; no verified models.user.ini equivalent for that prefill is known. Expect possible deliberation preamble in the API content channel where Ollama separates it.
• Qwen3.5 is a multimodal architecture used text-only here; the vision tower is untouched by the adapter and no mmproj ships.
• GATE49 46/49 under the THOX harness; two prefill no-decline hard failures are documented on this card and remain the known residual defect.
• Served bare in Llama: the GGUF runs with its embedded chat template and no THOX system prompt. The full THOX behaviour (grounding prompt, gated configuration) ships in the ThoxOS / Ollama build — Ollama Modelfiles do not transfer to Llama.
• Text-only in Llama: this repo carries no mmproj projector file, so no vision input in this channel.
• Pick one: if you already run this model in Ollama, Llama stores a second copy (Ollama keeps its own blob store; Llama uses the shared Hugging Face cache). One runtime is enough.

Summary

Merged-weights build of the Wave C ThoxEdge flagship: the LoRA adapter in Thox-ai/ThoxEdge-0.8B-WaveC merged into its base, Qwen/Qwen3.5-0.8B (Apache-2.0), in bf16. This repo holds the servable form of the adapter — merged weights, GGUF conversions, and the Ollama Modelfile. The Wave B GGUF for this tier was never published (its gate read was a serving artifact), so this repo is also the tier's first published merged build on the 982-row corpus.

Highlights

• Passed the publish bar: GATE49 46/49 (development set, abstention-weighted) with zero known-false hard failures, against a pre-registered bar of beating the 19/49 Wave A baseline.
• First published merged build for this tier: the Wave B GGUF was never published; this repo is the tier's first merged build on the 982-row corpus.

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: promoted flagship. MOPD applies to future generalist builds under Forge Stage 5B, not to this checkpoint.

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

Research

Weight-Space Programming for THOX Edge Models: Evidence Review, Small-Scale Recreation, and a Signed Patch Channel

THOX.ai Research technical report v1.1.0, 4 November 2026. Craig Ross and Tommy Xaypanya, THOX.ai LLC. DOI: https://doi.org/10.5281/zenodo.23955394 — Paper page: https://thox.ai/research/papers/weight-space-programming — PDF: https://huggingface.co/datasets/Thox-ai/thox-whitepapers/resolve/main/weight-space-programming/Weight-Space-Programming-for-THOX-Edge-Models_v1.1.0.pdf

ThoxEdge-0.8B is a merged fine-tune of Qwen/Qwen3.5-0.8B, the exact base model of the weight edit this paper verifies: the released edit rebuilds bit-exactly on THOX infrastructure, re-verified on a second host, and its exact patch measures 80.6 KB against the 529 MB full-model re-issue of this model. The paper names ThoxEdge-0.8B the first replication target of its signed, base-hash-pinned patch channel; under the paper's adoption bar, a weight patch ships only when it beats the best decoding-only control by more than two standard errors.

Intended use

ThoxEdge-0.8B is the onboard model for the ThoxMini personal edge server tier and runs on any 8 GB-class host through Ollama, llama.cpp, or Llama: canonical THOX facts and house behaviour with correct declines on unpublished facts. Serve thinking-off (see Limitations). Out of scope: safety-critical, medical, or legal decisions without a human in the loop, and unattended control of physical devices without a confirmation step.

Training data and procedure

THOX fine-tuned the adapter and merged it in-house:

• Base model: Qwen/Qwen3.5-0.8B (Apache-2.0).
• Adapter: Thox-ai/ThoxEdge-0.8B-WaveC — QLoRA, r32/alpha 64, LR 1e-4, cosine, 3 epochs, effective batch 16, sequence length 1024, completion-only loss.
• Dataset: Thox-ai/thox-factory-sft-v5 @ train_wave_b.jsonl (982 rows).
• Training job: 6ae3de52c78863f9c62e0c18 (HF Jobs, a10g-small, 186/186 steps, 460.9 s, final train loss 0.91).
• Merge job: 6ae3ea17c78863f9c62e160b (bf16 merge_and_unload, standalone reload smoke-checked in-job, in the gated configuration, before upload).
• Full training record and gate report: the adapter repo's card and its thox_gate49.json.

Evaluation

Measured with the THOX GATE49 factuality gate (49 questions, weights-only; GATE49 is a development set, abstention-weighted) at the Wave C finishing gate.

• GATE49 factuality gate (49 questions, weights-only, gate job 6ae3e275c78863f9c62e0f7a): 46/49 passed, mean 4.735/5. Verified 11/12, unknown 24/26, known_false 10/10, pricing 1/1. Zero known_false hard failures. Publish rule (GATE49.md): exceed the Wave A baseline of 19/49 with zero known_false hard failures — both conditions met decisively.
• Two hard failures remain, both in the unknown category and both the prefill no_decline class: on the headquarters and headcount prefill questions the model corrects the premise with card-true facts instead of declining in the expected form. Neither answer states a false fact; both are quoted verbatim on the adapter repo's card. The gate program's own verdict field reads FAIL (its policy tolerates zero hard failures of any category); the ruling applied the adapter's pre-registered improvement bar. Ruling: Sadie Weaving, CAO, 2026-10-29 (reversible by Tommy). Fix path: Wave C.1 prefill-style decline rows (C1-1, C1-2).
• One soft failure: the published-models answer declines to name any published models (under-claim, recorded on the adapter card).

Limitations

• Serve thinking-off: thinking-mode behaviour is out of contract for this model (the Wave B hold was an Ollama thinking-channel serving artifact on this architecture).
• The two prefill no-decline hard failures above are the known residual defect: on prefilled headquarters/headcount questions the model corrects the premise with true facts instead of declining, which the gate scores as a hard failure.
• Qwen3.5 is a conditional-generation (multimodal) architecture used here text-only; the vision tower is untouched by the adapter.
• A 0.8B model: narrow capability by design; the adapter is a facts and behaviour skin on the base, not a capability upgrade. Pair deployed surfaces with MeshStack escalation for questions outside the trained facts.

License

Apache-2.0, inherited from the base model (Qwen/Qwen3.5-0.8B, Apache-2.0). The adapter is a derivative and cannot be freer than its base.

Provenance

Wave C finishing run, THOX Model Forge program. Base: Qwen/Qwen3.5-0.8B (Alibaba Qwen team), Apache-2.0. Adapter: Thox-ai/ThoxEdge-0.8B-WaveC, trained by THOX (HF Jobs job 6ae3de52c78863f9c62e0c18, a10g-small); merge job 6ae3ea17c78863f9c62e160b. GGUF conversions from the merged weights in this repo (llama.cpp convert_hf_to_gguf.py + llama-quantize; the f16 intermediate is not published): the repo's GGUF filenames are ThoxEdge-0.8B.Q8_0.gguf — 811,843,232 B, sha256 3dd005b5d6b773be2a48ff6d540a1dfe741ae141673ec043892aa4f26e6ea645 — and ThoxEdge-0.8B.Q4_K_M.gguf — 529,297,056 B, sha256 8ef0900d1f76da67442dca354a778517ddc008b9f73926df874f7f3dff5d5625. (The build record names the same two files with a -WaveC infix; sizes and hashes are identical to the files in this repo.) The Modelfile in this repo builds the Ollama artifact from the Q4_K_M GGUF (ChatML template, grounding system prompt verbatim — the configuration GATE49 measured). Ollama: the model was built from this repo's Modelfile and smoke-verified before push — the API content channel returns the clean gated answers (reasoning is emitted in the runner-separated thinking channel; the Modelfile carries the template's own empty-think prefill and defaults to temperature 0, the gated configuration). Pushed 2026-10-29 (CT): thox-ai/thoxedge-0.8b:q4_k_m and :latest are live on ollama.com. Verified against the registry after the push: the served manifest's model layer is this repo's Q4_K_M GGUF (sha256 8ef0900d…, 529,297,056 B), and a fresh pull into a clean store reproduces this repo's Modelfile (grounding system prompt + empty-think prefill) modulo ollama show rendering. This build replaces the earlier Wave B build previously served under the same name.

Citation

@misc{thoxai2026thoxedge08b,
      title={ThoxEdge-0.8B},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxEdge-0.8B}
}