A THOX configuration and distribution of Mistral AI's Ministral 3 8B, the smallest and lightest model in the MagStack cluster family, published for local inference and MeshStack deployments. Apache-2.0.

Thox-Cluster-Swift — Speed-Oriented Configuration (Ministral 3 8B, THOX-Configured)

Your AI. Your Data. Your Rules.™

Thox-Cluster-Swift is a THOX configuration and distribution of Mistral AI’s Ministral 3 8B, the smallest and lightest model in the MagStack cluster family, published for local inference and MeshStack deployments. The weight layer in this model’s Ollama manifest is byte-identical to the upstream provider’s Ollama weight layer (ministral-3:8b). THOX’s work in this release is the MeshStack clustering layer these configurations were made for, configuration (runtime parameters and, where present, a system prompt), packaging, and distribution. THOX did not train these weights.

Correction notice. This page previously advertised specific tokens-per-second, first-token latency, concurrency, and uptime figures; those were inherited catalog copy with no THOX deployment measurement behind them and have been removed. No benchmark, throughput, latency, concurrency, uptime, cost-saving, or compliance figure is claimed for this release; figures of that kind that appeared on earlier versions of this page were inherited catalog copy and have been removed pending THOX-run evaluation on THOX deployments. Earlier versions of this page also stated the license as MIT; that was incorrect (see License).

Run it

Ollama (published model)

ollama pull Thox-ai/thox-cluster-swift
ollama run Thox-ai/thox-cluster-swift

Summary

A THOX configuration and distribution of Ministral 3 8B for local inference and MeshStack deployments: the smallest download in the cluster family (6.0 GB), suited to lighter edge hardware. Configuration-derived: THOX did not train these weights. This distribution’s weight layer is the same artifact as thox-pro-8b’s (identical SHA-256 digest); the two differ in configuration and presentation. Cluster Swift is packaged for multi-device team serving, Pro 8B for single-device use.

Specification	Details
Base model	Ministral 3 8B by Mistral AI
Parameters	8.9B (config model_type), mistral3 family, vision-capable (Text and Image per the manifest family record)
Quantization	Q4_K_M (GGUF file_type in the manifest config)
Context window (served)	32,768 tokens, from the params layer (num_ctx) in this model’s Ollama manifest. The Ollama page tag display of 256K is a display artifact and does not describe the served configuration.
Deployment	Local inference via Ollama; distributed orchestration via MeshStack (separate from this download)

Intended use

Local inference through Ollama, on hardware that fits an 8B Q4-class model:

Interactive local chat and assistance workloads on modest hardware
High-volume, short-query local inference where a small model is the right tool
Vision-plus-text tasks such as document and chart reading, within the base model’s abilities

Text and image input, per the upstream model’s capabilities. Downloading this model gives you local inference in Ollama. Distributed orchestration across devices is provided by MeshStack; a model download alone does not create a cluster or pool memory across devices.

Training data and procedure

No training by THOX. Base: mistralai/Ministral-3-8B-Instruct-2512 by Mistral AI, via Ollama tag ministral-3:8b. The original THOX release materials for this model claim no separate THOX fine-tuning run, training dataset, or training procedure, and none is claimed here. Base-model training data and procedure are those of the base model’s provider; see the base model card for that record. Datasets added by THOX for this release: none.

Evaluation

No THOX-run evaluation has been published for this model, and no benchmark scores were published on the original card. THOX claims no throughput, latency, or concurrency figure for this release; the performance figures on earlier versions of this page were removed as unmeasured (see the correction notice). This section will be updated with method and results when a THOX-run evaluation lands.

Hardware Reference (from the original card)

Reference configuration from the original card, per device: minimum 16 GB RAM, 15 GB storage, NVIDIA Jetson Orin Nano class GPU (8GB VRAM). Recommended 32 GB RAM, 25 GB NVMe SSD, Jetson Orin AGX class GPU.

Limitations

Outputs can be wrong. Professional, medical, legal, financial, and security-relevant outputs require qualified human review.
Local deployment only in this distribution; behavior and speed depend entirely on the operator’s hardware and configuration.
The operator is responsible for meeting any regulatory obligations that apply to their deployment.
A small model has the reasoning limits of its class; verify important outputs.
All processing is local by design; the model has no access to online information or external services unless the operator wires them in.

License

The model weights are distributed under the Apache License 2.0 (license layer in this model’s Ollama manifest, from the upstream ministral-3:8b distribution). Earlier versions of this page stated MIT; that was incorrect. THOX claims no license over the weights beyond the upstream terms.

MagStack and MeshStack technology is proprietary to Thox.ai LLC. Publication of this model does not license THOX mesh technology, patents, or trademarks beyond identifying the model.

Provenance

THOX MeshStack Cluster family. First listed on Ollama under Thox-ai in 2025 and held private; publicly released 2026-10-31 by order of Tommy Xaypanya, CTO and Co-Founder of THOX.ai, as part of the MeshStack model family release. Upstream: mistralai/Ministral-3-8B-Instruct-2512, Ollama tag ministral-3:8b; the weight layer is byte-identical to the upstream provider’s Ollama weight layer (SHA-256 verified in the 2026-10-31 release review) and the same artifact as thox-pro-8b’s weight layer. Model card: https://huggingface.co/Thox-ai/thox-cluster-swift. Correction history: inherited catalog performance figures (tokens-per-second, first-token latency, concurrency, uptime) removed as unmeasured and the license statement corrected from MIT to Apache-2.0, as recorded in the correction notice above. Enterprise: enterprise@thox.ai.