Thox-ai/README

THOX.ai — local-first AI, on hardware you control

Your AI. Your Data. Your Rules.™

THOX.ai builds local-first AI: purpose-built small models, personal devices, and the runtimes that serve them on hardware you own. No cloud dependence by default. This organization is the public home of our model releases; every repository carries its own model card, and the card is the authority for that artifact's files, license, and measured results.

As of November 4, 2026, this organization publishes 109 public models on Hugging Face and 89 models on Ollama, each with a full model card built to the THOX model page standard.

The devices. ThoxKey carries your models and workspace while a connected host does the compute. ThoxAir is the wireless edge node. ThoxMini is the personal edge server. ThoxClip clips onto your phone for onboard AI and wireless charging. The platforms are ThoxOS (the operating and intelligence layer), MeshStack (private connectivity and request routing between your devices), and ThoxWork (the agentic workspace). Lineup, kits, and delivery: thox.ai.

Flagship device models

Four models, one per device tier, trained in the THOX Wave program and scored on GATE49, our 49-question factuality gate. GATE49 is a development-set measure that rewards verified recall and careful declines; it is not a held-out benchmark, and no THOX model is qualified on physical device hardware yet. Full method and results: the Wave C whitepaper in the THOX whitepapers dataset.

Model	Tier	Status	GATE49
ThoxAir-0.5B	ThoxAir	Promoted flagship	47/49
ThoxEdge-0.8B	ThoxMini (onboard tier)	Promoted	46/49
ThoxMini-Phi4Mini	ThoxMini	Documented baseline (not promoted; see card)	47/49
ThoxClip-Qwen3-0.6B	ThoxClip	Published baseline, grounding required	22/49 (see card)

ollama run Thox-ai/thoxair-0.5b
ollama run Thox-ai/thoxedge-0.8b
ollama run Thox-ai/thoxmini-phi4mini
ollama run Thox-ai/thoxclip-qwen3-0.6b

Notes that matter: ThoxEdge-0.8B works best with thinking disabled ("think": false). ThoxClip-Qwen3-0.6B must be served with its grounding system prompt; its card explains why.

Host and laptop assistants (THOX-trained)

• ThoxForge-7B — code-oriented 7B assistant, LoRA adapter + bundled Q4_K_M GGUF. ollama run Thox-ai/thoxforge-7b (The ThoxForge-7B-GGUF repository is an intentional placeholder with no weights; the runnable build is in ThoxForge-7B.)
• ThoxGlobal-7B — multilingual 7B assistant, LoRA adapter + Q4_K_M GGUF. ollama run Thox-ai/thoxglobal-7b
• ThoxNova-12B-Core — conservative chat 12B, LoRA adapter + GGUF. ollama run Thox-ai/thoxnova-12b-core
• ThoxNova-12B-Agent — tool-calling agent 12B, LoRA adapter + GGUF. ollama run Thox-ai/thoxnova-12b-agent
• ThoxWave-8B — creative all-rounder, THOX QLoRA fine-tune (Llama 3 license). ollama run Thox-ai/thoxwave-8b
• ThoxMini-3B — lightweight edge assistant (Llama 3.2 license). ollama run Thox-ai/thoxmini-3b

ThoxOpen — configured open models

ThoxOpen-8B is the everyday laptop model of the THOX Open family: upstream Qwen3-8B weights packaged by THOX with the THOX Open system profile and the THOX Knowledge Pack. It is configuration-derived; THOX did not train these weights. The Hugging Face repository holds the packaging (Modelfile, manifest, knowledge pack); the runnable model is on Ollama:

ollama run Thox-ai/thoxopen-8b

The family's minimal-refusal build is ThoxOpen-8B-Heretic, THOX's own abliteration of Qwen3-8B, measured at 13/100 refusals against the stock model's 100/100: ollama run Thox-ai/thoxopen-8b-heretic.

MeshStack Cluster and Pro — configuration distributions

Ten models (Cluster: code, deep, maverick, scout, secure, swift; Pro: 8b, 14b, 24b, 32b) published as THOX configurations of upstream open models for MeshStack workgroups. The honest framing, stated on every card: the weight layers are byte-identical to upstream; THOX's work is the MeshStack clustering layer, configuration, and documentation. MeshStack routes whole requests between nodes; it does not split a model across devices. The Hugging Face repositories are the model cards; the runnable builds are on Ollama. THOX evaluation for this family is pending and each card says so. Two members (scout, secure) build on a non-commercial research-license base and carry that warning on their cards; they are not offered for commercial use.

Small-model and research lines

• Device tiers: ThoxEdge-0.5B, ThoxFlux-350M, ThoxNano-135M, ThoxNano-360M — earlier-wave models for the same device slots, each with its gate record on its card.
• Role models: the -role checkpoints (ThoxKey-9M-role, ThoxAir-16M-role, ThoxClip-9M-role, ThoxMini-125M-role) are small-model role research for device firmware tiers, published with their limits stated.
• From-scratch research: ThoxLLM, ThoxMicro, ThoxNeedle-Micro, and the TinyStories base are THOX's own small-model research — documented, and not shippable assistants.
• Routing and edge components: ThoxRoute-1.5B (model-routing classifier, with GGUF builds), ThoxMesh-Head-C6 (TFLite mesh event predictor), ThoxSpeech-ASR-BitNet, ThoxUSB-Edge.
• Browser tier: ThoxWebby-Gemma-4-E2B is a card-only repository for the ThoxOS browser tier; it distributes no weights, and the linked Gemma license governs the upstream model.

In production

ThoxOS-Kernel-27B — the ThoxOS host-tier kernel model: a full-parameter fine-tune, GATE-PASSED under the adopted final-answer scope (GATE49 41/49, development set), with weights released publicly on November 1, 2026 under the D11 open-weights ruling. Its students ship alongside it: ThoxOS-Kernel-4B, the resident kernel model for ThoxMini and ThoxClip (documented baseline, GATE49 44/49), and ThoxOS-Kernel-0.8B, the triage student for ThoxAir and ThoxKey (GATE-PASSED, GATE49 37/49). All three run on Ollama.

Uncensored redistributions — read before use

Four repositories redistribute community uncensored models. They are not trained by THOX and not safety-tuned; refusal behaviour is deliberately reduced by their upstream authors. They are public and ungated by deliberate founder decision. They are not part of the ThoxOS or MeshStack catalog and are not promoted for THOX devices. Apply your own safety layer before any user-facing use.

• ThoxHeretic-27B — abliterated 27B redistribution (Hugging Face only).
• ThoxHeretic-9B — abliterated Qwen3.5-9B merge with a paired vision projector; runnable build on Ollama: ollama run Thox-ai/thoxheretic:9b
• ThoxMythos-9B — uncensored community fine-tune redistribution.
• ThoxIntel-27B — uncensored 27B redistribution.

Licensing

Licensing is per repository and per artifact. Read the model card and the license in the repo you download; do not infer a license from a model family or from public availability. Most flagship derivatives inherit permissive base licenses (Apache-2.0 or MIT); Llama- and Gemma-based artifacts carry their upstream terms, which govern. Where a card and a footer disagree, the card's license section and the upstream terms are the authority.

Build with us

Model issue? Use that repo's Community tab with the model revision, file or quantization, runtime version, and a minimal reproduction. Never include credentials or private data.

Start from the THOX edge device models collection rather than the full repo list.

THOX.ai LLC — Your AI. Your Data. Your Rules.™

Organization overview updated November 4, 2026. Model cards and release notes carry artifact-specific status.