Explore THOX models for local AI, from compact device assistants to coding, multilingual chat and tool-calling models.
Public on Ollama and Hugging Face
Run these published models locally on a compatible computer without buying tokens or paying per prompt. No THOX device purchase or Kickstarter backing is required. Your computer supplies the memory and compute; each model has its own context and runtime limits.
The device names identify their intended roles. These exact builds are not yet qualified on physical THOX devices. ThoxKey uses the connected host computer for assistant inference.
Availability and model cards checked against the THOX Model Factory catalog, revision 2026-11-04.1 (reviewed 2026-11-04). This page lists all 94 published models in that catalog.
Compact Qwen- and Phi-based fine-tunes focused on THOX knowledge and assistance. Status and scores below are reported by THOX; read the evaluation notes and each card before choosing a build.
ThoxAir assistant
A compact THOX product and device-help model, fine-tuned from Qwen2.5-0.5B-Instruct. The current release is the Wave C.2 build.
Use the supplied grounding prompt. Narrow factual coverage; the card documents remaining errors.
ollama run Thox-ai/thoxair-0.5bThoxMini edge tier
A THOX product and tool-routing model, fine-tuned from Qwen3.5-0.8B. This release uses the model for text generation.
Serve with thinking disabled and the supplied grounding prompt. The card records two remaining hard failures.
ollama run Thox-ai/thoxedge-0.8b --think=falseThoxMini / ThoxKey host assistant
A Phi-4-mini-instruct fine-tune for THOX assistance and host-side serving. With ThoxKey, the connected host computer runs the assistant.
Grounding required. The deployment probe found spec, arithmetic and tool-routing failures; the card also records a funding fabrication.
ollama run Thox-ai/thoxmini-phi4miniThoxClip assistant baseline
A Qwen3-0.6B fine-tune for THOX product questions and device help. Factual reliability depends on serving it with reference facts.
Serve thinking-off with the grounded THOX device-facts system prompt. Bare deployment is outside the published serving requirements.
ollama run Thox-ai/thoxclip-qwen3-0.6b --think=falseThoxAir / ThoxKey kernel and OS triage
A Qwen3.5-0.8B LoRA fine-tune, merged, distilled from ThoxOS-Kernel-27B for kernel/OS triage questions on the smallest THOX devices.
Serve with the supplied grounding prompt. Known limits stay on the card: GATE-K verified recall 0/12, probes 6/20.
ollama run Thox-ai/thoxos-kernel-0.8bThoxMini / ThoxClip resident kernel and OS agent (baseline)
A Qwen3.5-4B LoRA fine-tune, merged, distilled from ThoxOS-Kernel-27B. Published as a documented baseline: its safety-form bars are held and named on the card.
Do not deploy as a resident OS agent without a harness that enforces destructive-action confirmation.
ollama run Thox-ai/thoxos-kernel-4bLarger assistants for compatible computers. These are public on both platforms. Hugging Face provides LoRA adapters that need their base models, plus standalone GGUF builds. No THOX GATE49 results or quality figures are published for the Forge, Global or Nova models; the kernel teacher’s gate record is summarized on its card above and in full on its model card.
Code-oriented assistant
A Mistral-7B-Instruct-v0.3 LoRA fine-tune for coding and general text generation on a compatible host.
LoRA adapter + Q4_K_M GGUF
ollama run Thox-ai/thoxforge-7b:Q4_K_MMultilingual assistant
A Qwen2.5-7B-Instruct LoRA fine-tune for multilingual text generation. Quality varies by language.
LoRA adapter + Q4_K_M GGUF
ollama run Thox-ai/thoxglobal-7b:Q4_K_MConservative chat
A Gemma-4-12B-it LoRA fine-tune with a chat persona. Core and Agent share the same THOX adapter with different persona overlays.
LoRA adapter + Q4_K_M / Q5_K_M GGUF
ollama run Thox-ai/thoxnova-12b-core:Q4_K_MTool-calling agent
The Nova agent persona for tool-calling workloads on a compatible host. Review and validate generated tool calls before executing them.
LoRA adapter + Q4_K_M / Q5_K_M GGUF
ollama run Thox-ai/thoxnova-12b-agent:Q4_K_MThoxOS kernel/OS model (host tier) and family teacher
A full-parameter fine-tune of the Qwen3.8-27B community build (JonathanColetti/Qwen3.8-27B-Uncensored) for Linux kernel and system operations, ThoxOS diagnostics, and MeshStack operations. Open weights under the D11 ruling; user-swappable in the ThoxOS kernel-model slot.
Safetensors (with MTP head file); Q8_0 and Q4_K_M GGUF
ollama run Thox-ai/thoxos-kernel-27b:Q4_K_MNova needs a runtime with support for its model family. Check the model card for hardware guidance, exact files and license terms. Nova here is a model family, not a current THOX device.
Every published model in the THOX Model Factory catalog (revision 2026-11-04.1) appears on this page: the device and host models above carry detailed cards, and the remaining 83 published models are listed here by family. Configuration-derived models state what THOX actually did in their notes; weights that are byte-identical to an upstream release say so. Build recipes, enterprise profiles and services are not downloadable models and are listed separately in the THOX Model Factory.
Wave B fp-LoRA. GATE49 47/49. Development-set, abstention-rewarding gate; not device-qualified.
ollama run Thox-ai/thoxedge-0.5bWave B LoRA. GATE49 20/49: a weak result, stated on its own listing.
ollama run Thox-ai/thoxflux-350mSmolLM-135M LoRA, ThoxKey tier. GATE49 23/49: a weak result, stated on its own listing.
ollama run Thox-ai/thoxnano-135mWave B LoRA. GATE49 23/49: a weak result, stated on its own listing.
ollama run Thox-ai/thoxnano-360mFrom-scratch 327M foundation model (v2), trained by THOX. Apache-2.0.
ollama run Thox-ai/thoxllmFrom-scratch ThoxMicro-125M base/completion model, trained by THOX.
ollama run Thox-ai/thoxmicroFrom-scratch BitNet b1.58 ternary decoder (9M parameters) trained by THOX on TinyStories. GGUF builds (Q8_0 and TQ2_0) for llama.cpp; Ollama and LM Studio load only the Q8_0 build. Base model for the ThoxClip-9M-Role and ThoxKey-9M-Role checkpoints.
From-scratch BitNet b1.58 ternary decoder (16M parameters) trained by THOX on TinyStories. GGUF builds (Q8_0 and TQ2_0) for llama.cpp; Ollama and LM Studio load only the Q8_0 build. Base model for the ThoxAir-16M-Role checkpoint.
Full fine-tune of ThoxMicro-125M for device-role tasks.
ollama run Thox-ai/thoxmini-125m-roleBitNet ternary 9M role checkpoint. Its Ollama listing states Ollama cannot load the ternary weights; run it with llama.cpp.
ollama run Thox-ai/thoxclip-9m-roleBitNet b1.58 ternary 16M role checkpoint, fine-tuned from ThoxMicro-1bit-16M (a THOX model trained from scratch). Apache-2.0.
ollama run Thox-ai/thoxair-16m-roleBitNet ternary 8.9M firmware role model for ThoxKey.
ollama run Thox-ai/thoxkey-9m-rolePublished 2026-11-01. ThoxKey micro assistant (host-assisted configuration). THOX-configured Qwen2.5-0.5B, not THOX-trained. 34.08 tok/s long-form, 0.59 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Apache-2.0.
ollama run Thox-ai/thoxkey-micro-0.5bPublished 2026-11-01. ThoxKey routing-proposal configuration. THOX-configured Qwen2.5-0.5B, not THOX-trained. 33.67 tok/s long-form, 0.59 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Apache-2.0.
ollama run Thox-ai/thoxkey-route-0.5bPublished 2026-11-01. ThoxAir routing configuration (2 GB target). THOX-configured Qwen2.5-0.5B, not THOX-trained. 55.35 tok/s long-form, 0.59 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Apache-2.0.
ollama run Thox-ai/thoxair-route-0.5bPublished 2026-11-01. ThoxAir chat configuration (2 GB target). THOX-configured Qwen2.5-0.5B, not THOX-trained. 58.8 tok/s long-form, 0.59 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Apache-2.0.
ollama run Thox-ai/thoxair-chat-0.5bPublished 2026-11-01. iOS micro model for intent classification. THOX-configured Qwen2.5-0.5B, not THOX-trained. 61.97 tok/s long-form, 0.59 GB RAM on CPU at 8,192 ctx; grounding 5/8, trap declined. Apache-2.0.
ollama run Thox-ai/thox-ios-micro-0.5bPublished 2026-11-01. ThoxMini decision configuration (4 GB target). THOX-configured Qwen2.5-1.5B, not THOX-trained. 11.35 tok/s long-form, 1.37 GB RAM on CPU at 8,192 ctx; grounding 7/8, trap declined. Apache-2.0.
ollama run Thox-ai/thoxmini-mind-1.5bPublished 2026-11-01. ThoxClip decision configuration (6 GB target). THOX-configured Qwen2.5-1.5B, not THOX-trained. 20.87 tok/s long-form, 1.37 GB RAM on CPU at 8,192 ctx; grounding 7/8, trap declined. Apache-2.0.
ollama run Thox-ai/thoxclip-mind-1.5bPublished 2026-11-01. iOS edge assistant and routing model. THOX-configured Qwen2.5-1.5B, not THOX-trained. 35.67 tok/s long-form, 1.37 GB RAM on CPU at 8,192 ctx; grounding 7/8, trap declined. Apache-2.0.
ollama run Thox-ai/thox-ios-edge-1.5bPublished 2026-11-01. ThoxMini chat configuration (4 GB target). THOX-configured Qwen2.5-3B, not THOX-trained. 11.66 tok/s long-form, 2.42 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Qwen Research License (non-commercial). The base model is under the Qwen Research License (non-commercial).
ollama run Thox-ai/thoxmini-chat-3bPublished 2026-11-01. ThoxClip chat configuration (6 GB target). THOX-configured Qwen2.5-3B, not THOX-trained. 11.03 tok/s long-form, 2.42 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Qwen Research License (non-commercial). The base model is under the Qwen Research License (non-commercial).
ollama run Thox-ai/thoxclip-chat-3bPublished 2026-11-01. iOS grounded chat model. THOX-configured Qwen2.5-3B, not THOX-trained. 11.32 tok/s long-form, 2.42 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Qwen Research License (non-commercial). The base model is under the Qwen Research License (non-commercial).
ollama run Thox-ai/thox-ios-chat-3bPublished 2026-11-01. THOX iOS system profile over Qwen3-1.7B (stand-in: Apple Foundation Models weights are not distributed). THOX-configured, not THOX-trained. 24.37 tok/s long-form, 2.43 GB RAM on CPU at 8,192 ctx; grounding 6/8, trap declined. Apache-2.0.
ollama run Thox-ai/thox-ios-systemPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-4B (Apache-2.0): ThoxClip target agent profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 15.28 tok/s long-form, 3.89 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined.
ollama run Thox-ai/thoxclip-agent-4bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-4B (Apache-2.0): THOX iOS Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 15.34 tok/s long-form, 3.89 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined.
ollama run Thox-ai/thox-ios-open-4bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-8B (Apache-2.0): THOX iOS Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 8.38 tok/s long-form, 6.68 GB footprint at 8,192 context (CPU), factory 3/3, grounding 7/8 (one topic's run was lost when the host OOM-killed the inference runner mid-pass; the other seven passed and the earlier pass scored 8/8), fabricated-topology trap declined after profile strengthening.
ollama run Thox-ai/thox-ios-mesh-8bDolphin-2.9 Llama-3-8B QLoRA fine-tune. Public and ungated (2026-10-31); the Llama 3 license travels with the model and governs use.
ollama run Thox-ai/thoxwave-8bLlama-3.2-3B QLoRA fine-tune. Public and ungated (2026-10-31); the Llama 3.2 license governs use. Distinct from the ThoxMini-Phi4Mini flagship.
ollama run Thox-ai/thoxmini-3bThoxRoute-1.5B routing classifier. Published evaluation on its Ollama page: 289/289 on a held-out template split and 7/8 on a hostile out-of-distribution probe.
ollama run Thox-ai/thoxrouteRelease Set 2, published 2026-10-31. THOX configuration of Qwen2.5-Coder-7B (Apache-2.0): Code system profile, Knowledge Pack, and serving parameters; the weights are the upstream model's. Rebased to the dedicated Qwen2.5-Coder 7B base in the audited registry.
ollama run Thox-ai/thoxcode-8bRust-leaning fine-tune trained by THOX on a public Apache-2.0 Rust code corpus (74,606 filtered code-producing examples): rank-32 LoRA merged into the base weights. MultiPL-E HumanEval-Rust 28 on the card (a same-harness delta measurement). Serves cheap Rust completion in ThoxRoute's lowest tier; a hosted/desktop model whose GGUF ladder spans about 260 MB, so it does not fit THOX's smallest devices. Hugging Face only; no Ollama release.
Published 2026-11-01 on Hugging Face and Ollama. THOX Edge laptop configuration of Qwen3-4B (Apache-2.0): system profile from the THOX Model Release Kit plus THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Not the trained flagship ThoxEdge-0.8B. Measured on the card: 15.19 tok/s long-form, 3.89 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined.
ollama run Thox-ai/thoxedge-4bPublished 2026-11-01. Agent Routing & Classification Model (Qwen3-0.6B, THOX-Configured) — thoxagent-route-0.6b is the routing tier of the THOX Agent family: upstream Qwen3-0.6B weights packaged with a THOX agent system profile specialized for classification and
ollama run Thox-ai/thoxagent-route-0.6bPublished 2026-11-01. Local Tool-Plan Drafting Model (Qwen3-1.7B, THOX-Configured) — thoxagent-edge-1.7b is the edge tier of the THOX Agent family: upstream Qwen3-1.7B weights packaged with a THOX agent system profile specialized for local tool-plan drafting, pl
ollama run Thox-ai/thoxagent-edge-1.7bPublished 2026-11-01 on Hugging Face and Ollama. Core tier of the THOX Agent family: THOX configuration of Qwen3-4B (Apache-2.0) with the THOX Agent DigitalHumans kit agent profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 15.12 tok/s long-form, 3.89 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined. Method note from the card: Qwen3-4B thinking mode consumed the full token cap on several first attempts, so those prompts were rerun with thinking disabled and the reruns scored.
ollama run Thox-ai/thoxagent-core-4bTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-cluster-codeTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-cluster-deepTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING. Qwen Research License: non-commercial use only. No certification or authorization is claimed for this configuration.
ollama run Thox-ai/thox-cluster-secureTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING. Qwen Research License: non-commercial use only.
ollama run Thox-ai/thox-cluster-scoutTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-cluster-swiftTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-cluster-maverickTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-pro-32bTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-pro-24bTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-pro-14bTHOX configuration of upstream weights: the weight layers are byte-identical to upstream; THOX work is the MeshStack configuration layer, documentation, and origin program. The Hugging Face repo is the model card; the runnable build is on Ollama. THOX evaluation is PENDING.
ollama run Thox-ai/thox-pro-8bRelease Set 1, published 2026-10-31. THOX configuration of Qwen3-30B-A3B (Apache-2.0): Mesh system profile, Knowledge Pack, and serving parameters; the weights are the upstream model's. MeshStack routes whole requests to a node; no distributed or split inference is claimed.
ollama run Thox-ai/thoxmesh-30b-a3bPublished 2026-11-01. THOX configuration of Qwen3-30B-A3B (Apache-2.0): agent system profile, enterprise reasoning prompt pack, THOX Knowledge Pack v0.1, and serving parameters; the weights are the upstream model's. Measured 52.23 tok/s long-form on a workstation host, 2026-11-01. MeshStack routes whole requests to a node; no distributed or split inference is claimed.
ollama run Thox-ai/thoxmesh-agent-30bPublished 2026-11-01. THOX configuration of Qwen3-30B-A3B (Apache-2.0): MeshStack expansion system profile, THOX Knowledge Pack v0.1, and serving parameters; the weights are the upstream model's. Measured 53.75 tok/s long-form on a workstation host, 2026-11-01. MeshStack routes whole requests to a node; no distributed or split inference is claimed.
ollama run Thox-ai/thoxmesh-moe-30bPublished 2026-11-01. THOX configuration of Qwen3-30B-A3B (Apache-2.0): enterprise workgroup system profile, workgroup prompt pack, THOX Knowledge Pack v0.1, and serving parameters; the weights are the upstream model's. Measured 53.68 tok/s long-form on a workstation host, 2026-11-01. Workgroup isolation is a prompting convention, not a security boundary. MeshStack routes whole requests to a node; no distributed or split inference is claimed.
ollama run Thox-ai/thoxmeshgroup-30b-a3bPublished 2026-11-01. MeshStack dense baseline on Qwen3-32B with the THOX Knowledge Pack. 4.87 tok/s long-form, 22.61 GB footprint on a GPU host at 8,192 ctx; factory 3/3, grounding 8/8, trap declined. Config-only, not THOX-trained. Apache-2.0.
ollama run Thox-ai/thoxmesh-32bPublished 2026-11-01. MeshStack routing assistant on Qwen3-14B with the THOX Knowledge Pack. 25.25 tok/s long-form, 10.49 GB footprint on a GPU host at 8,192 ctx; grounding 8/8, trap declined. Config-only, not THOX-trained. Apache-2.0.
ollama run Thox-ai/thoxmesh-open-14bPublished 2026-11-01. MeshStack routing assistant on dense Qwen3-32B with the THOX Knowledge Pack. 5.39 tok/s long-form, 22.61 GB footprint on a GPU host at 8,192 ctx; grounding 8/8, trap declined. Config-only, not THOX-trained. Apache-2.0.
ollama run Thox-ai/thoxmesh-dense-32bPublished 2026-11-01. Workgroup assistant on dense Qwen3-32B with THOX profiles and Knowledge Pack. 5.36 tok/s long-form, 22.61 GB footprint on a GPU host at 8,192 ctx; grounding 8/8, trap declined. Config-only, not THOX-trained. Apache-2.0.
ollama run Thox-ai/thoxmeshgroup-32bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-8B (Apache-2.0): THOX Work MeshGroup Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Federated means site-level request routing only; no distributed or split inference is claimed. Measured on the card: 8.75 tok/s long-form, 6.68 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined.
ollama run Thox-ai/thoxmeshgroup-federated-8bRelease Set 1, published 2026-10-31. THOX configuration of Qwen3-8B (Apache-2.0): system profile and serving parameters; the weights are the upstream model's.
ollama run Thox-ai/thoxopen-8bPublic per founder ruling 2026-10-31. Community reduced-refusal build redistributed by THOX; not THOX-trained and no THOX evaluation is claimed.
ollama run Thox-ai/thoxheretic:9bPublic per founder ruling 2026-10-31. Community redistribution; not THOX-trained and no THOX evaluation is claimed.
ollama run Thox-ai/thoxmythosPublic per founder ruling 2026-10-31. Community Qwen3.6-27B merge build, not THOX-trained: the Hugging Face repo is card-only and the runnable build is on Ollama (slug thoxintel). Vision probe passed 2026-11-01; THOX evaluation pending.
ollama run Thox-ai/thoxintelFully public and ungated per founder ruling 2026-10-31. Hugging Face only: it is not distributed through Ollama. Not THOX-trained; a GGUF redistribution whose card lists SHA-256 hashes and states no THOX evaluation exists.
Card and documentation only: this repository distributes no weights. Public and ungated (2026-10-31); the linked Gemma license governs the upstream model.
Release Set 2, published 2026-10-31. THOX configuration of Qwen3-4B (Apache-2.0): Open system profile, Knowledge Pack, and serving parameters; the weights are the upstream model's.
ollama run Thox-ai/thoxopen-4bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-14B (Apache-2.0): THOX Open system profile and THOX Knowledge Pack v0.1; the weights are the upstream base model weights, pulled by Ollama from the pinned base, and THOX did not train them. Measured on the card: 1.04 tok/s long-form (cloud CPU), 10.76 GB loaded at 8,192 context, factory 3/3, grounding 8/8, routing PASS, fabricated-topology trap declined.
ollama run Thox-ai/thoxopen-14bPublished 2026-11-01. Minimal-refusal vision model: THOX Open profile on abliterated Gemma 4 E2B (QAT Q4), projector included. 120.44 tok/s, 1.86 GB footprint on a GPU host at 8,192 ctx; grounding 8/8, trap declined. Upstream weights. Apache-2.0.
ollama run Thox-ai/thoxopen-vision-hereticPublished 2026-11-01. THOX's own abliteration build from stock Qwen3-8B (Apache-2.0), produced with the Heretic tool (heretic-llm 1.4.0). Standalone re-evaluation of the shipped weights: refusals 13/100 against 100/100 stock, KL divergence 0.0482, capability spot-check 8/8. Downloadable weights on Hugging Face (abliterated safetensors plus Q4_K_M, Q5_K_M and Q8_0 GGUF builds; the Q8_0 build was added 2026-11-02 and passed its per-quant gate: refusal probes 0/25, benign 0/40, spot-check 8/8) and live on Ollama (tags latest, q5_k_m, q4_k_m).
ollama run Thox-ai/thoxopen-8b-hereticPublished 2026-11-01. Persona-following reasoning model on Qwen3-14B with THOX profiles and Knowledge Pack. 27.67 tok/s long-form, 10.49 GB footprint on a GPU host at 8,192 ctx; grounding 8/8, trap declined. Config-only, not THOX-trained. Apache-2.0.
ollama run Thox-ai/thoxhuman-reason-14bPublished 2026-11-01. THOX-configured Qwen3-8B model for digital-human conversation, with the THOX agent profile and Knowledge Pack v0.1. 49.78 tok/s long-form, 6.31 GB footprint (GPU host, 2026-11-01). Config-derived; THOX did not train these weights. Apache-2.0.
ollama run Thox-ai/thoxhuman-chat-8bPublished 2026-11-01. THOX digital-human persona: Sadie Weaving, Chief Agentic Officer. 9.13 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-sadiePublished 2026-11-01. THOX digital-human persona: Ava, Primary DigitalHuman, platforms and fleet orchestration. 8.99 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-avaPublished 2026-11-01. THOX digital-human persona: THOXY, THOX platform guide. 8.34 tok/s, 3.89 GB at 8,192 ctx. Configuration-derived from Qwen3-4B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-thoxyPublished 2026-11-01. THOX digital-human persona: Mira, Knowledge SME. 13.87 tok/s, 3.89 GB at 8,192 ctx. Configuration-derived from Qwen3-4B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-miraPublished 2026-11-01. THOX digital-human persona: Kai, Device worker. 13.78 tok/s, 3.89 GB at 8,192 ctx. Configuration-derived from Qwen3-4B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-kaiPublished 2026-11-01. THOX digital-human persona: Sera, Compliance judge. 8.11 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-seraPublished 2026-11-01. THOX digital-human persona: Nova, Coding and build. 8.91 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-novaPublished 2026-11-01. THOX digital-human persona: Helix, Finance SME. 9.14 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-helixPublished 2026-11-01. THOX digital-human persona: Rhea, People manager. 8.85 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-rheaPublished 2026-11-01. THOX digital-human persona: Atlas, Strategy architect. 8.67 tok/s, 6.68 GB at 8,192 ctx. Configuration-derived from Qwen3-8B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-atlasPublished 2026-11-01. THOX digital-human persona: Ginger Weaving, Chief Innovation & Engineering Officer. 13.75 tok/s, 3.89 GB at 8,192 ctx. Configuration-derived from Qwen3-4B; not THOX-trained. Apache-2.0.
ollama run Thox-ai/thox-gingerPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-4B (Apache-2.0): THOX Work MeshGroup Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 15.32 tok/s long-form, 3.89 GB footprint at 8,192 context (CPU), factory 3/3, grounding 7/8 (the routing-policy topic failed), fabricated-topology trap declined. Upgraded from the earlier enterprise-profile recipe entry now that the model itself is published.
ollama run Thox-ai/thoxwork-office-4bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-4B (Apache-2.0): THOX Work MeshGroup Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 15.32 tok/s long-form, 3.89 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined. Upgraded from the earlier enterprise-profile recipe entry now that the model itself is published.
ollama run Thox-ai/thoxwork-mind-4bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-8B (Apache-2.0): THOX Work MeshGroup Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 8.73 tok/s long-form, 6.68 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined. Upgraded from the earlier enterprise-profile recipe entry now that the model itself is published.
ollama run Thox-ai/thoxwork-analyst-8bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-8B (Apache-2.0): THOX Work MeshGroup Model Kit profile and THOX Knowledge Pack v0.1; the weights are byte-identical to upstream and THOX did not train them. Measured on the card: 8.59 tok/s long-form, 6.68 GB footprint at 8,192 context (CPU), factory 3/3, grounding 8/8, fabricated-topology trap declined on pass 2 after profile strengthening (the first build played along with the trap; only the strengthened build was published). Upgraded from the earlier enterprise-profile recipe entry now that the model itself is published.
ollama run Thox-ai/thoxwork-agent-8bPublished 2026-11-01 on Hugging Face and Ollama. THOX configuration of Qwen3-14B (Apache-2.0): THOX Work coding system profile, enterprise workgroup prompt pack and THOX Knowledge Pack v0.1; the weights are the upstream base model weights and THOX did not train them. Measured on the card: 26.28 tok/s long-form (GPU workstation, shared host, a floor), 10.49 GB footprint at 8,192 context, factory 3/3, grounding 8/8, fabricated-topology trap declined.
ollama run Thox-ai/thoxwork-code-14bGATE49 is THOX’s internal 49-question factuality check: verified THOX facts, questions the model should decline, false premises it should challenge, and pricing. It is not a general intelligence benchmark or evidence of performance on your hardware. The scores above summarize producer-reported evaluations, not independent runtime tests or proof that every quantization behaves identically.
Each linked model card carries the full method, known defects and serving requirements. Grounding supplies reference facts; it does not guarantee a correct answer.
GGUF is a packaged model format commonly used by local runtimes; quantization reduces file size and can change quality. Available quantizations differ by model. Download size is not the same as memory needed to run it: context length, runtime and hardware all affect memory use and speed.
With local inference, prompts are processed on your computer. Connected tools, mesh nodes and remote providers have their own data paths; review them before sharing private information.
Your AI. Your Data. Your Rules.™
This documentation is provided for informational and operational purposes only. The specifications and technical details herein are subject to change without notice. THOX.ai LLC reserves all rights in the technologies, methods, and implementations described.
Nothing in this documentation shall be construed as granting any license or right to use any patent, trademark, trade secret, or other intellectual property right of THOX.ai LLC, except as expressly provided in a written agreement.
The MagStack™ magnetic stacking interface technology is proprietary technology of THOX.ai LLC, protected by trade secrets and intellectual property laws....
Except as permitted by applicable law or a component’s applicable license, you may not reverse engineer, disassemble, decompile, decode, or otherwise attempt to derive the source c...
THOX.ai™, ThoxOS™, MagStack™, MeshStack™, ThoxMigrate™, the THOX Edge Series™, the THOX Nova Series™, and the THOX.ai logo are trademarks of THOX.ai LLC. WireGuard® is a registered trademark of Jason A. Donenfeld.
All other trademarks are the property of their respective owners.
© 2026 THOX.ai LLC. All Rights Reserved.