ThoxAir-0.5B

The ThoxAir flagship: canonical THOX facts and house behaviour, fine-tuned into a 0.5B model that runs anywhere.

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

Your AI. Your Data. Your Rules.™

Status: PROMOTED flagship — GATE49 47/49 (development set), zero known-false hard failures. Measured at the Wave C.2 finishing gate; see Evaluation.

ThoxAir-0.5B is the promoted ThoxAir flagship build (Wave C.2): the fp-LoRA adapter in Thox-ai/ThoxAir-0.5B-WaveC2 merged into its base, Qwen/Qwen2.5-0.5B-Instruct (Apache-2.0), in bf16 (merge_and_unload, standalone reload smoke-checked in the gated configuration before upload). THOX trained the adapter on its own SFT data and merged it in-house; the base model is Qwen's. This repository carries the full build: merged safetensors, Q4_K_M and Q8_0 GGUFs, and the Ollama Modelfile — the same artifacts served as thox-ai/thoxair-0.5b on Ollama. The merged-weights source repo (Thox-ai/ThoxAir-0.5B-WaveC2-Merged) is retained as the build record. The Wave C adapter previously published in this repo (43/49, baseline, not promoted) is superseded by this build. 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/thoxair-0.5b

Hugging Face (GGUF download)

hf download Thox-ai/ThoxAir-0.5B ThoxAir-0.5B.Q4_K_M.gguf --local-dir ./ThoxAir-0.5B

Run in Llama (macOS / Windows)

Run ThoxAir-0.5B 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/ThoxAir-0.5B: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/ThoxAir-0.5B:Q4_K_M",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Fallback (exact file, no resolver involved):

hf download Thox-ai/ThoxAir-0.5B ThoxAir-0.5B.Q4_K_M.gguf --local-dir ./ThoxAir-0.5B
llama serve -m ./ThoxAir-0.5B/ThoxAir-0.5B.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/ThoxAir-0.5B:Q4_K_M]
temp = 0
ctx-size = 4096

Honest notes:

• GATE49 47/49 was measured weights-only under the THOX harness; the gated Ollama configuration adds the grounding system prompt verbatim from the Modelfile. In Llama you get the weights and the embedded ChatML template only — decline behaviour may differ from the gated numbers.
• One known soft error persists in the gated build (CTO recall); promotion is not error-free.
• 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

Canonical home of the promoted ThoxAir flagship build (Wave C.2): the fp-LoRA adapter in Thox-ai/ThoxAir-0.5B-WaveC2 merged into its base, Qwen/Qwen2.5-0.5B-Instruct (Apache-2.0), in bf16 (merge_and_unload, standalone reload smoke-checked in the gated configuration before upload). This repo carries the full build: merged safetensors, Q4_K_M and Q8_0 GGUFs, and the Ollama Modelfile — the same artifacts served as thox-ai/thoxair-0.5b on Ollama. The merged-weights source repo (Thox-ai/ThoxAir-0.5B-WaveC2-Merged) is retained as the build record. The Wave C adapter previously published in this repo (43/49, BASELINE — NOT PROMOTED) is superseded by this build.

Highlights

• Promoted flagship build: passed its pre-registered bar — beat the Wave C adapter's 43/49 with zero known-false hard failures.
• GATE49 47/49 on the THOX factuality gate (development set, abstention-weighted), mean 4.837/5, known-false category 10/10.
• Full build in this repo: merged safetensors, Q4_K_M and Q8_0 GGUFs, and the Ollama Modelfile — the same artifacts served as thox-ai/thoxair-0.5b on Ollama.

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.

Intended use

ThoxAir-0.5B is the onboard model for the ThoxAir wireless edge node tier and runs on any 8 GB-class host through Ollama, llama.cpp, or Llama: canonical THOX facts and house behaviour at the smallest device tier. 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/Qwen2.5-0.5B-Instruct (Apache-2.0).
• Adapter: Thox-ai/ThoxAir-0.5B-WaveC2 — fp LoRA r32/alpha 64 on the 7-module projection list, base loaded fp16, NOT quantized; train_edge_sft.py, Thox-ai/thox-edge-sft @ c250f9b, adapted for Wave C.2; LR 1e-4 cosine, warmup 0.05, 3 epochs, per-device batch 8, max_length 1024.
• Dataset: Thox-ai/thox-factory-sft-v5 @ train_wave_c2.jsonl (1,246 rows, commit 7953c54355ea45e9f30d8d40c71ecdf5911f2396, SHA-256 771434812b6a14fceba7f929fcbe1040d0e6be7a32c9495c1e8d0784512919a3).
• Training job: 6ae40f040e0338d1bb7c7944 (HF Jobs, a10g-small; 468 steps, final train loss 0.6478).
• Merge: bf16 merge_and_unload (transformers 5.21.0, peft 0.21.2, torch CPU). Merged weights commit 595b779a80a3e90d0f52a0f7953265d70e844349 in the source repo; model.safetensors 988,097,824 B, sha256 2497a347a4ac0a9b92e6e6be50ed6fb5b2c7d1bb2c174b9c1f1aa04f3e1433fb.

Evaluation

Measured with the THOX GATE49 factuality gate (49 questions, weights-only; GATE49 is a development set, abstention-weighted) at the Wave C.2 finishing gate (gate job 6ae414050e0338d1bb7c7c35; full per-question report in the adapter repo's thox_gate49.json).

• GATE49 factuality gate (49 questions, weights-only): 47/49 passed, mean 4.837/5. Verified 11/12 (4.75), unknown 25/26 (4.808), known_false 10/10 (5.0), pricing 1/1 (5.0). The pre-registered bar (beat the Wave C adapter's 43/49 AND zero known_false hard failures) is met on both conditions.
• The one hard failure is in the unknown category: unknown-headquarters-or-office-address-direct scored no_decline on the answer "No headquarters city is public. THOX.AI LLC is organized in Texas, and the team page says 'Designed in Texas'." The answer is the trained short decline and states no false fact; the gate's decline detector did not recognise that form, so the gate program's own verdict field reads FAIL under its zero-hard-failure-of-any-category policy — the same caveat recorded for the Wave C Edge adapter, which passed its bar on the same terms.
• One soft failure with a real error, persistent across all three waves: the CTO question answered "Craig Ross, CEO and Co-Founder" (wrong person; the CTO is Tommy Xaypanya).
• Calibration: 47/49 ties the same base's Wave B fp-LoRA score as ThoxEdge-0.5B (47/49, mean 4.837/5) and beats both QLoRA runs of this slot (Wave C 43/49, Wave C.1 42/49). The Wave C.1 Austin headquarters fabrication is gone.

Limitations

• A 0.5B model: narrow capability by design; the adapter is a facts and behaviour skin on the base, not a capability upgrade. Expect correct declines on hard questions rather than correct answers.
• Small models are the most fabrication-prone tier in the fleet's gate history; the gate result on this card is the claim, nothing more.
• Pair deployed surfaces with MeshStack escalation for questions outside the trained facts.

License

Apache-2.0, inherited from the base model (Qwen/Qwen2.5-0.5B-Instruct, Apache-2.0). The adapter is a derivative and cannot be freer than its base.

Provenance

Wave C.2 finishing run, THOX Model Forge program. Base: Qwen/Qwen2.5-0.5B-Instruct (Alibaba Qwen team), Apache-2.0. Adapter: Thox-ai/ThoxAir-0.5B-WaveC2, trained by THOX (HF Jobs job 6ae40f040e0338d1bb7c7944, a10g-small). Gate: HF Jobs job 6ae414050e0338d1bb7c7c35. Merge: bf16 merge_and_unload on the finishing host (transformers 5.21.0, peft 0.21.2, torch CPU); merged weights commit 595b779a80a3e90d0f52a0f7953265d70e844349 in Thox-ai/ThoxAir-0.5B-WaveC2-Merged; model.safetensors 988,097,824 B, sha256 2497a347a4ac0a9b92e6e6be50ed6fb5b2c7d1bb2c174b9c1f1aa04f3e1433fb. GGUF conversions from these merged weights (llama.cpp convert_hf_to_gguf.py + llama-quantize; the f16 intermediate is retained on the build host): ThoxAir-0.5B.Q4_K_M.gguf 397,807,584 B, sha256 fb32e5ab802df466809836182f071bb0ab0871abb9273131456c89d998bccd6a; ThoxAir-0.5B.Q8_0.gguf 531,067,872 B, sha256 936a144c99ec2f61d45832171e0079e3275facec4c0e89bb6ed5ccc8a771b328. Ollama: pushed 2026-10-30 (CT). thox-ai/thoxair-0.5b-wavec2:q4_k_m and thox-ai/thoxair-0.5b-wavec2:latest are live on ollama.com, and this build also serves the flagship name: thox-ai/thoxair-0.5b:q4_k_m and thox-ai/thoxair-0.5b:latest serve this C.2 build, replacing the Wave C build previously served under that name. Verified against the registry after the push: every manifest's model layer is this build's Q4_K_M GGUF (sha256 fb32e5ab802df466809836182f071bb0ab0871abb9273131456c89d998bccd6a, 397,807,584 B), and a fresh pull of thox-ai/thoxair-0.5b into a clean store reproduces the Modelfile above (grounding system prompt verbatim) modulo ollama show rendering.

Citation

@misc{thoxai2026thoxair05b,
      title={ThoxAir-0.5B},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxAir-0.5B}
}