THOX.ai

Edge AI computing devices and from-scratch language models for local, privacy-preserving inference. No cloud. No data uploads. Your weights, your hardware, your network.

Profile: https://huggingface.co/Thox-ai Website: https://www.thox.ai GitHub: https://github.com/ttracx Org owner: @tommytracx License default: Apache-2.0 (open source); proprietary firmware kept separate.

What we ship

THOX.ai owns the stack end to end. Three layers, all of them real today.

1. Foundation models trained from scratch on THOX-defined architectures. Not fine-tunes of someone else's base.
2. ThoxOS kernel, runtime, and userspace orchestration framework for the device fleet.
3. Edge devices the models run on. ThoxClip, ThoxMini, ThoxNova, ThoxMicro, and the MagStack Air cluster variants.

Models on this org

Repos are private during the 2026-Q3 hardening window. Email tommy@thox.ai or open an issue at https://github.com/ttracx/thoxllm for access.

Foundation training roadmap

ThoxLLM-1B (target: 2026-06 to 2026-07). 24 layers, hidden 2048, GQA 32Q/8KV, ctx 4096, vocab 50304, bf16. 200k steps on wikitext-103 with grad-accum 64. Two architecture shapes scaffolded; final pick lands 2026-06-09:

The 327M result (val ppl 22.03 at 50k) is the floor; the 1B run is 4x params at 4x compute and should converge below 12 ppl on the same data at 200k steps.

ThoxLLM-3B (target: 2026-Q3). 32 layers, hidden 3072, GQA 24Q/8KV. Architecture identical to 1B; trained once the 1B run is published.

ThoxLLM-7B / 13B (target: 2026-Q4). Match Llama 3 dimensions. Multi-node training on a clustered MagStack Air rig.

Open-source layer

These ship on GitHub as Apache-2.0, not on HF. Linked here because the models target them.

Device fleet (current and committed)

The August 2026 Kickstarter ships ThoxClip ($39+), ThoxMini ($69), ThoxAir ($79), ThoxNova ($499). Goal: $250K baseline / $3M ceiling.

Open known gaps

Honest list of what's not done.

1. GGUF for ThoxLLM-327M is blocked. llama.cpp expects three FFN tensors per layer (gate / up / down). The current THOX SwiGLU has two (fc_in / fc_out). Either we (a) add a LLM_ARCH_THOX arch + kernel to a llama.cpp fork (2-3 days) or (b) switch ThoxLLM-1B+ to standard two-projection SwiGLU. Decision lands 2026-06-09; Path B is the recommended call in docs/decisions/2026-06-07-swiglu-shape-decision.md.
2. ThoxLLM-327M is not converged. Val loss was still descending at 50k. The 1B run targets 200k+ steps for a real converged baseline.
3. Repos are private. Hardening window ends 2026-Q3; expect public re-publication of ThoxLLM-327M and ThoxLLM-1B after the SwiGLU decision is committed.
4. No HF Inference API endpoint for the THOX base architecture yet. Single-projection THOX SwiGLU requires the thoxllm package; standard-SwiGLU 1B+ will work in vanilla Transformers.

Quickstart (once a model goes public)

from huggingface_hub import snapshot_download
from thoxllm import ThoxLLM, ThoxTokenizer

# Download via huggingface-cli login + access grant
local = snapshot_download("Thox-ai/ThoxLLM-327M", token=HF_TOKEN)
tok = ThoxTokenizer.from_pretrained(local)
model = ThoxLLM.from_pretrained(local, dtype="bfloat16", device_map="auto")

out = model.generate(tok("Edge AI means", return_tensors="pt").to(model.device),
                    max_new_tokens=64)
print(tok.decode(out[0], skip_special_tokens=True))

For the fine-tuned adapter family (Forge / Mini / Global / Wave / Nova), a GGUF Q4_K_M is shipped alongside the safetensors and runs in Ollama directly:

ollama create thox-forge-7b -f Modelfile.thox-forge-7b
ollama run thox-forge-7b "Explain the THOX edge architecture in two paragraphs."

Compliance and posture

THOX devices run a stripped Linux-compatible userspace on top of ThoxOS Kernel. The whole pipeline is auditable:

• All training data is documented in repo dataset.md files.
• No telemetry from devices. ThoxOS Kernel has zero outbound network calls in the boot path.
• Compliance gates (HIPAA / GDPR / SOC2 / PCI / ISO27001) are enforced at the kernel syscall layer via thox-compliance. See ADR-0006 in the kernel repo.
• Persona binding (THOXY / Mira / Sera) constrains what each model context can read or write per device session.

Roadmap (only what is committed)

Contact

• Tommy Xaypanya (@tommytracx) - founder and engineer of record
• tommy@thox.ai for org member requests and partnership inquiries
• Issues + discussion: https://github.com/ttracx/thoxllm and https://github.com/ttracx/thoxos-kernel

---

Updated 2026-06-07. Last 30 days: shipped ThoxLLM-327M trained run, six fine-tune adapter pairs, ThoxOS Kernel v0.1.0 -> v1.0.0 (18 ADRs), THOX DigitalHumans v0.1.0, THOX Experience Fabric v0.3.0, ThoxLiteRT-LM, ThoxMesh, Thox Portable v0.2.0, Thox Terminal v0.1.0, MagStack Air family (3 repos), ThoxMicro.