Skip to content

Cluster Models

THOX cluster distributions, their upstream model provenance, and how they differ from the current device and host assistants.

Model releases and cluster serving

A model artifact and a distributed runtime solve different problems. The new THOX assistants can be tried on a compatible host computer. Making a model available to authorized devices on a network is different from splitting one inference job across multiple computers.

A public cluster-model listing is not evidence that any particular group of THOX devices can run it. Device count alone does not establish pooled memory, context capacity, speed or concurrent-user capacity. Exact hardware, runtime and network qualification remain necessary.

Public cards, links and Ollama manifests checked October 31, 2026. No cluster inference was run for this page.

Where the new models fit

Device-model program

These are compact THOX fine-tunes with model-specific grounding requirements. Their GATE49 results are factuality results, not distributed-runtime or physical-device qualification. ThoxKey uses host compute.

Host assistants

Start with the published GGUF on a compatible host, or load an adapter with its required base model. These releases do not need a cluster merely because they are part of the THOX catalog. Network serving depends on your configured runtime and permissions.

Current model downloads and commandsDevice and host compatibility

Published cluster distributions

The seven listings below are public THOX-configured distributions of upstream models. Their current cards describe THOX configuration, documentation and the MeshStack / MagStack program; they do not claim a separate THOX weight-training run. This differs from the THOX fine-tunes in the Model Catalog.

Hugging Face hosts the model cards for these seven listings, not their model weights. Runnable artifacts are listed on Ollama. The cards carry upstream license obligations and distinguish original catalog figures from THOX-run evaluations. No benchmark, speed, uptime, device-count or workload-certification claim is made here.

Entry configuration

Cluster Small

Thox-ai/thox-cluster-small

Upstream base named by the card
NVIDIA Nemotron 3 Nano 30B
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Code-oriented configuration

Cluster Code

Thox-ai/thox-cluster-code

Upstream base named by the card
Qwen2.5-Coder-32B-Instruct
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Interactive serving configuration

Cluster Swift

Thox-ai/thox-cluster-swift

Upstream base named by the card
Ministral-3-8B-Instruct-2512
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Large-model reasoning configuration

Cluster Deep

Thox-ai/thox-cluster-deep

Upstream base named by the card
Llama-3.1-405B-Instruct
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Isolated-deployment configuration

Cluster Secure

Thox-ai/thox-cluster-secure

Upstream base named by the card
Qwen2.5-72B-Instruct
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Department serving configuration

Cluster Scout

Thox-ai/thox-cluster-scout

Upstream base named by the card
Qwen2.5-72B-Instruct
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Large MoE configuration

Cluster Maverick

Thox-ai/thox-cluster-maverick

Upstream base named by the card
Qwen3-235B-A22B
Publication
Public card on Hugging Face; artifact on Ollama
Qualification
Check the exact runtime and deployment; availability is not cluster validation
Ollama listing Provenance and license

Cluster Small is the current public listing for the entry configuration. Do not substitute the older thox-cluster-nano or coordinator examples as install commands; those public download names were not available in this check. Scout and Maverick use the Qwen bases named above, rather than the Llama 4 bases described by the older website table.

Validate your serving configuration

  1. Choose the exact model, quantization and runtime from its linked card. Check the upstream license and available artifact before planning hardware.
  2. Confirm that the runtime can load the model on the selected host or supports the specific distribution method you need. An Ollama model name alone does not configure cross-device inference.
  3. For a distributed job, document the runtime build, per-node memory, interconnect, model partitioning and context settings. Measure latency and throughput with your actual workload.
  4. Check which authorized node receives prompts and which tools or remote providers are connected. Local model weights do not define every data path in a mesh.
Start with a local modelMeshStack overviewMeshStack privacy model

CONFIDENTIAL AND PROPRIETARY INFORMATION

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.

Patent Protection

The MagStack™ magnetic stacking interface technology is proprietary technology of THOX.ai LLC, protected by trade secrets and intellectual property laws....

Reverse Engineering Prohibited

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.