ThoxKey-9M-role — canonical ThoxKey device model

Device-lock note (2026-10-23): under the current THOX device lock, ThoxAir and ThoxClip do not host a language model, and ThoxMini on-device model support is held until a test pass is recorded. Role models are trained for these device personas and run on a paired host (phone, laptop or MeshStack node). Throughput claims were removed pending a recorded measurement. Owner wording required before merge.

Capability status (2026-09-29): Public weights. Device mapping is recorded in ttracx/thox-device-models and certified only as public + mapped + no blocked marketing claims. This is not a tok/s, AIME, or HIPAA certificate. ThoxNova names on this org are a model family, not a campaign SKU.

This is the model baked into the ESP32-S3 firmware. 8,917,248 params, BitNet b1.58 ternary ({-1, 0, +1} weights, int8 activations — W1.58A8).

Lineage — verified, not assumed

Fine-tuned from Thox-ai/ThoxMicro-1bit-9M, a THOX model trained from scratch on TinyStories. There is no external base — the ternary pretrain is ours.

Precisely: training initialised from the 9M pretrain checkpoint (micro-9m-ternary/best.pt, 393M tokens), and the ThoxMicro-1bit-9M repo publishes that same run as GGUF. The repo holds no .pt, which is why the init_from in configs/ternary/thoxkey-9m-role.yaml points at the local checkpoint rather than the Hub path — same weights, different container.

A previous internal proposal listed slvDev/esp32-ai as a base. That was wrong and is retracted. It is a GitHub repository (github.com/slvDev/esp32-ai, MIT), not a Hugging Face model, so it could never be an HF base_model — and no THOX model was trained from it. The shipping ThoxKey model is this one.

What it does

Handles local device commands, names itself when asked, and defers everything else rather than inventing it. It holds no company facts by design — at 8.9M params (23.5% of it the embedding) there is no capacity for a fact table, and training one in is how larger models in this family learned to fabricate.

prompt	answer
who are you	I am ThoxKey, a THOX model running locally on the key.
do you send my data anywhere	No. I run on the key. Nothing leaves it.
who funded thox	(defers — no company facts in the corpus)

Zero fabrications on held-out company-fact prompts.

Files

file	use
thoxkey-9m-role.tern1	the firmware artifact — TERN1, ESP32-S3 micro runtime
thoxkey-9m-role-TQ1_0.gguf	llama.cpp ternary
thoxkey-9m-role-Q8_0.gguf	llama.cpp / Ollama host validation

Ollama cannot load ternary GGUF (TQ1_0/TQ2_0/I2_S) — tensor size overflow. Use Q8_0 on a host; TQ1_0/TERN1 are the device artifacts.

Device budget: 3.80 MB weights, 8.00 MB resident (throughput figure removed pending a recorded on-target measurement). ESP32-S3 N16R8: fits.

Ship-then-test: not measured on target silicon. Trained on a local RTX 4060 Ti at $0.

Temporary demo Space: tommytracx/ThoxKey-9M-Space (org Space paused on quota).

Files

File	Size	SHA-256
hf/model.safetensors	17.84 MB (17,842,448 B)	9762cf4d55a5333029c7eddf955d8c23287edb540071446b5cbc9cdf6cea4b18
thoxkey-9m-role-Q8_0.gguf	9.77 MB (9,771,104 B)	c63fbde3efae89061e71c1a779c78d7c28f7cfb11c66458ffd7a54c0aec807d2
thoxkey-9m-role-TQ1_0.gguf	5.93 MB (5,933,152 B)	d3725928dc1579bd0f6b7b9484b8bd5615b8dcf5e0384091952b29e481374f98
thoxkey-9m-role-TQ2_0.gguf	6.25 MB (6,252,640 B)	3ba414c4cd2939d22c3ee8be3f9105e10a80c1047ac9a9cc2c849af53c3da2e4
thoxkey-9m-role.tern1	4.42 MB (4,420,028 B)	741f06a01cb1369f87f97d02b1f562c2ac3a24a24ecd4d36247541acacfc15b6

Use with ThoxOS

Download the file once and run it fully offline on a ThoxOS host (or any Linux/macOS/Windows machine):

pip install -U "huggingface_hub[cli]"
hf download Thox-ai/ThoxKey-9M-role thoxkey-9m-role-Q8_0.gguf --local-dir ./ThoxKey-9M-role
llama-cli -m ./ThoxKey-9M-role/thoxkey-9m-role-Q8_0.gguf -p "Hello" -n 128

Ollama (pulls straight from the Hub):

ollama run hf.co/Thox-ai/ThoxKey-9M-role:Q8_0

Ternary TQ1_0 / TQ2_0 files need a recent llama.cpp build; Ollama and LM Studio may only load the Q8_0 file.

LM Studio: search Thox-ai/ThoxKey-9M-role in the model browser (or lms get Thox-ai/ThoxKey-9M-role), pick thoxkey-9m-role-Q8_0.gguf, and load it.

Use with MeshStack

Serve the model as an OpenAI-compatible endpoint on any MeshStack node, then register that endpoint with your MeshStack coordinator (protocol openai, base URL http://<node-ip>:8080/v1; or protocol ollama at http://<node-ip>:11434 if you serve it with Ollama).

llama-server -m ./ThoxKey-9M-role/thoxkey-9m-role-Q8_0.gguf --host 0.0.0.0 --port 8080 -c 512
curl http://<node-ip>:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Hello"}]}'

Use the prompt format documented above; small/role models may need -c 512 and a plain completion prompt.

Device fit (by file size only)

• thoxkey-9m-role-TQ1_0.gguf (5.93 MB): Tiny (<50 MB): runs on any host; also small enough for constrained edge targets that have a GGUF-capable runtime.
• thoxkey-9m-role-TQ2_0.gguf (6.25 MB): Tiny (<50 MB): runs on any host; also small enough for constrained edge targets that have a GGUF-capable runtime.
• thoxkey-9m-role-Q8_0.gguf (9.77 MB): Tiny (<50 MB): runs on any host; also small enough for constrained edge targets that have a GGUF-capable runtime.

RAM needed ≈ file size + context (KV cache) + runtime overhead. These tiers are derived from file size alone; no throughput, TOPS or battery figures are claimed here.

Licence and attribution

Licence pending review: trained on TinyStories (CDLA-Sharing-1.0, share-alike). Treat redistribution terms as unresolved until THOX confirms; do not bundle commercially in the meantime.

Out of scope

• Medical, legal, financial, safety-critical or other high-stakes decisions without human review.
• Unattended control of physical devices or actuators without a confirmation step.
• Any use prohibited by the upstream licence or acceptable-use policy.

Limitations and bias

• Small and quantized models make factual mistakes and can produce confident but wrong output; verify important answers.
• Behaviour and bias reflect the training data, which is mostly English web/instruction text; quality in other languages is lower unless stated above.
• Quantization trades accuracy for size; lower-bit files degrade more.

---

THOX.ai LLC — Your AI. Your Data. Your Rules.™ · thox.ai · Contact: tommy@thox.ai