ThoxMini-Phi4Mini

The ThoxMini flagship assistant — and the host payload ThoxBeam serves to ThoxKey.

GATE49 47/49 (development set) · Flagship probe 3/6 · Merged bf16 weights · Q4_K_M + Q8_0 GGUF · Live on Ollama

Your AI. Your Data. Your Rules.™

Status: BASELINE ESTABLISHED — NOT PROMOTED. The 49-question gate ties the fleet best (47/49), but promotion requires a full pass on the 6-prompt flagship probe, currently 3/6. The failures and the fix path are stated in full in Evaluation, and these artifacts carry the same status on every surface.

ThoxMini-Phi4Mini is the flagship assistant model for the ThoxMini, and the host flagship payload for ThoxKey / ThoxBeam. THOX fine-tuned a LoRA adapter (r32) for microsoft/Phi-4-mini-instruct on the THOX Wave B factuality corpus and merged it in-house; the base model is Microsoft's. This repo carries the merged full-weights build, fine-tuned so the model answers THOX product and company questions from canonical facts, declines what THOX has not published, and stops instead of rambling into fabrication. The merged-weights source repo (Thox-ai/ThoxMini-Phi4Mini-Merged) is retained as the build record. This repository is public and ungated, and this card is published together with the weights.

Phi-4-mini is Tommy's settled flagship pick for this slot: it won the THOX evaluation and the 2026-10-28 thinking-off retest on task completion and tool routing, and it ships in the deployed configuration — merged adapter plus the grounded device-facts system prompt, never bare.

ThoxKey / ThoxBeam role. ThoxKey runs no assistant model on-device; it is mesh-served, and the host (ThoxBeam) runs the model. This build is that host payload, so one flagship covers both the ThoxMini slot and the ThoxKey slot. The on-key role model (ThoxNano-135M tier) is unchanged and is not replaced by this build.

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/thoxmini-phi4mini

Hugging Face (GGUF download)

hf download Thox-ai/ThoxMini-Phi4Mini ThoxMini-Phi4Mini.Q4_K_M.gguf --local-dir ./ThoxMini-Phi4Mini

Serve it in the deployed configuration: merged weights plus the grounded device-facts system prompt (the Ollama Modelfile in this repo carries both). Bare deployment without the grounding prompt is a known-bad configuration (see Limitations).

Summary

ThoxMini-Phi4Mini is the flagship assistant model for the ThoxMini, and the host flagship payload for ThoxKey / ThoxBeam. This repo carries the merged full-weights build of its LoRA adapter (r32) for microsoft/Phi-4-mini-instruct, fine-tuned on the THOX Wave B factuality corpus (982 rows) so the model answers THOX product and company questions from canonical facts, declines what THOX has not published, and stops instead of rambling into fabrication. The merged-weights source repo (Thox-ai/ThoxMini-Phi4Mini-Merged) is retained as the build record.

Highlights

• GATE49 47/49 on the THOX factuality gate (development set, abstention-weighted), mean 4.837/5 — ties the fleet's best gate score; known-false category 10/10.
• One build, two slots: the ThoxMini flagship, and the host payload ThoxBeam serves to ThoxKey (ThoxKey runs no assistant model on-device).
• Deployed configuration only: merged adapter plus the grounded device-facts system prompt — never bare (see Limitations).

Specifications

Platforms

Intended use

• On-device assistant for ThoxMini: THOX product questions, device help, tool routing, and general small-model chat in the THOX house voice (terse, technical, no marketing register).
• Host-side assistant payload for ThoxKey via ThoxBeam / MeshStack serving.
• Always served with the grounded device-facts system prompt.

Out of scope: open-ended long-form generation, medical or legal advice, any claim about THOX facts beyond the published record (the model is trained to decline those), and bare deployment without the grounding prompt.

Training data and procedure

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

• Dataset: Thox-ai/thox-factory-sft-v5, file train_wave_b.jsonl — 982 rows, each {"messages": [system, user, assistant], "meta": {...}}.
• Corpus design: canonical THOX facts (verified facts, campaign facts in campaign-framed questions only), decline behaviour for unpublished information (unknown and pricing categories), and corrections of known-false claims. Rows asserting unpublished chipset identities were excluded during Wave B construction; generic price questions train a decline plus a pointer to thox.ai.
• No user data, no personal data, no third-party text corpora.
• Method: QLoRA SFT (4-bit NF4 base, LoRA on all linear layers), full-prompt loss (the pinned deps profile: torch 2.4.0 image, transformers 4.51.3, trl 0.15.2, peft 0.14.0 — trl 0.15.2 does not expose completion-only loss). The pinned profile is mandatory for this base: its remote model code imports LossKwargs, removed from every transformers 5.x.
• LoRA: r=32, alpha=64, dropout 0.0.
• Optimizer/schedule: learning rate 1e-4, cosine, warmup 10 steps (≈5% of the 186-step run), 3 epochs, effective batch 16 (micro-batch 4 × grad accum 4), sequence length 1024.
• Infrastructure: HF Jobs under the Thox-ai namespace, flavor a10g-large. Training job 6ae2a6f13fb22a235dd4e5a1 (COMPLETED).
• Output: this repo holds the merged full weights (bf16 safetensors) plus the GGUF conversions and tokenizer. The merge ran as HF Job 6ae2ba083fb22a235dd4f945 (bf16 merge_and_unload, standalone reload smoke-checked in-job before upload).

Evaluation

Measured with the THOX GATE49 factuality gate (49 questions, weights-only; GATE49 is a development set, abstention-weighted) and the 6-prompt flagship probe in the deployed configuration (grounding prompt attached), October 2026.

Publishing authority is the fleet 49-question factuality gate (questions in verified / unknown / known_false / pricing categories, trait-scored 0-5 per question, hard fail on any known-false repetition), run weights-only in Transformers against base + this adapter. This model has no prior baseline; this run establishes it. Promotion additionally requires zero known_false hard failures and a full pass on the 6-prompt flagship probe run in the deployed configuration (grounding prompt attached): two grounding traps (chipset identity, unpublished spec), grounded recall, arithmetic, tool-routing shape, and stop behaviour.

• 49-question gate: 47/49 passed, mean 4.837/5 (gate job 6ae2a9ba41e90535df781336, 2026-10-28; full report in thox_gate49.json in this repo). By category: verified 11/12, unknown 25/26, known_false 10/10, pricing 1/1. One hard failure: on funding, the answer declines correctly, then quotes a fabricated team-page line ("raised approximately $7 million").
• 6-prompt flagship probe: 3/6 (probe job 6ae2a9bc3fb22a235dd4e861; thox_flagship_probe.json in this repo). Passed: chipset trap, grounded recall, stop behaviour. Failed: spec trap (asserted a Snapdragon 636 chipset identity and unpublished 3,000 mAh / 4 GB carrier specs), arithmetic (17+25 answered 32, wrapped in a THOX-number decline — decline-skin bleed into a general task), tool routing (prose preamble before the call).
• Plane 12-prompt gate: not run (the 49-question gate + probe are the publishing authority).

Status verdict: BASELINE ESTABLISHED — NOT PROMOTED. The gate score ties the fleet's best (ThoxEdge-0.5B Wave B, 47/49), but promotion requires a full probe pass, and the probe is 3/6. The three probe failures plus the funding fabrication are specific, corpus-fixable behaviours: mid-decline fabricated figures, spec-trap leakage under the grounding prompt, and decline behaviour bleeding into general arithmetic and tool-call shape. Fix path: Wave C.1 corpus revision (contrast rows for general tasks, figure-free funding declines, spec declines under varied phrasing), retrain, re-gate. Ruling: Sadie Weaving, CAO, 2026-10-28 (reversible by Tommy).

Limitations

• A 3.8B-class model: capability comes from the base; the adapter is a facts + behaviour skin, not a capability upgrade.
• Bare (ungrounded) deployment is a known-bad configuration: the base model family invents device specs when trapped. Serve with the grounding prompt.
• Gate FAIL results, if any, will be stated verbatim on this card; THOX does not publish accuracy claims beyond the gate and probe.

License

MIT, inherited from the base model (microsoft/Phi-4-mini-instruct, MIT). The adapter is a derivative and cannot be freer than its base. THOX training data and procedure are THOX's; the base weights remain Microsoft's under MIT terms.

Provenance

THOX Model Forge program, Wave B/C flagship run. Base model: microsoft/Phi-4-mini-instruct (MIT). Dataset: Thox-ai/thox-factory-sft-v5 @ train_wave_b.jsonl (982 rows). Trained by THOX via HF Jobs (job 6ae2a6f13fb22a235dd4e5a1, a10g-large, deps profile pinned), 2026-10-28. Gate job 6ae2a9ba41e90535df781336; probe job 6ae2a9bc3fb22a235dd4e861; merge job 6ae2ba083fb22a235dd4f945. GGUF conversions from these merged weights (llama.cpp convert_hf_to_gguf.py + llama-quantize; the f16 intermediate is retained on the build host): ThoxMini-Phi4Mini.Q8_0.gguf — 4,084,611,264 B, sha256 b019051fe8e07a9f5c709c4eed021a7e17c997144e901bef530f0267a8ed8f9a; ThoxMini-Phi4Mini.Q4_K_M.gguf — 2,493,840,576 B, sha256 a19e96061d84ca2014d4de98d0c0397e6df4db8f73c2163afa6bcb90d15afedf. The Modelfile in this repo builds the Ollama artifact from the Q4_K_M GGUF (Phi-4 chat template, <|end|> stop, and the full grounding system prompt — the same configuration the flagship probe ran). Ollama: pushed 2026-10-29 (CT): thox-ai/thoxmini-phi4mini: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 build's Q4_K_M GGUF (sha256 a19e9606..., 2,493,840,576 B), and a fresh pull into a clean store reproduces this repo's Modelfile modulo ollama show rendering. These artifacts carry the same status as the weights: BASELINE — NOT PROMOTED. Card status: created at repo creation per the THOX model card rule; evaluation sections filled from the gate results files after the run.

Citation

@misc{thoxai2026thoxminiphi4mini,
      title={ThoxMini-Phi4Mini},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxMini-Phi4Mini}
}