Model Compatibility
Which THOX device or runtime can run each model: ThoxKey + ThoxBeam, ThoxAir, ThoxMini, ThoxClip, host runtimes, and MeshStack.
Last updated 9 months ago
All model entries below are sourced from the published model-compat manifest. Verified means evidence exists for that model/runtime pairing; it does not imply that every THOX device can run the model locally.
Manifest-driven model compatibility
This page answers one question: where can a model actually run? It is generated from the published compatibility manifest and refreshed every five minutes. Device role models, ThoxBeam aliases, host-class models, and MeshStack workloads are kept distinct so model availability is not confused with device capability.
Vision Models
Native image understanding, medical imaging, OCR in 32+ languages.
Multilingual
Support for 12-32+ languages including English, Spanish, Chinese, Arabic.
Performance
Quantization and runtime choice determine fit. Check the current device/runtime row instead of assuming a model is local to every THOX device.
Verified Compatibility
Models with published compatibility evidence. Always check the device/runtime fit before treating a verified model as locally runnable.
Filter Models
Coordinator
Alibaba · Qwen 3
Lightweight cluster orchestration and management model.
Size
4B
Speed
<100ms
Memory
3GB
Context
16K tokens
Min Devices
1x
Evidence
Jetson AI Lab
Device Fit
THOX-ai/thox-cluster-coordinatorMinistral-3 8B
Mistral · Ministral-3
Edge-optimized vision model with 32+ languages. Perfect for single devices with vision needs.
Size
8B
Speed
40-60 tok/s
Memory
10GB
Context
256K tokens
Min Devices
1x
Evidence
Jetson AI Lab
Device Fit
Best For
ministral-3:8bGemma 3 8B
Google · Gemma 3
Google's efficient vision model optimized for single GPU. Excellent balance of performance and capability.
Size
8B
Speed
38-55 tok/s
Memory
10GB
Context
128K tokens
Min Devices
1x
Evidence
Benchmark
Device Fit
Best For
gemma3:8bQwen 3 14B
Alibaba · Qwen 3
Advanced reasoning model with vision and multilingual support. Excellent for complex professional tasks.
Size
14B
Speed
30-45 tok/s
Memory
14GB
Context
128K tokens
Min Devices
1x
Evidence
Jetson AI Lab
Device Fit
Best For
qwen3:14bPhi-4 Mini (3.8B)
Microsoft · Phi-4 Mini
Microsoft's compact model with exceptional performance. Multilingual with function calling.
Size
3.8B
Speed
70-95 tok/s
Memory
4GB
Context
128K tokens
Min Devices
1x
Evidence
Benchmark
Device Fit
Best For
phi4:miniLlama 3.2 8B
Meta · Llama 3.2
Meta's reliable foundation model. Excellent for general professional use.
Size
8B
Speed
42-65 tok/s
Memory
10GB
Context
128K tokens
Min Devices
1x
Evidence
Vendor Docs
Device Fit
Best For
llama3.2:8bQwen 2.5 Coder 14B
Alibaba · Qwen 2.5 Coder
State-of-the-art coding model with reasoning improvements and 128K context.
Size
14B
Speed
28-42 tok/s
Memory
14GB
Context
128K tokens
Min Devices
1x
Evidence
Benchmark
Device Fit
Best For
qwen2.5-coder:14bDeepSeek-Coder-V2 16B
DeepSeek · DeepSeek-Coder-V2
Advanced coding model with MoE architecture. Excellent for software engineering.
Size
16B
Speed
25-38 tok/s
Memory
16GB
Context
64K tokens
Min Devices
1x
Evidence
Benchmark
Device Fit
Best For
deepseek-coder-v2:16bCompatibility Watchlist
Unverified — under evaluation. Not yet recommended for production workloads.
Cluster 70B
Alibaba · Qwen 3
Performance benchmarks pending Jetson AI Lab validation.
- Size
- 72B
- Memory
- 140GB
- Min Devices
- 2x
- Status
- Review required
THOX-ai/thox-cluster-70bCluster 100B
Alibaba · Qwen 3
Memory footprint requires 4x stack; quantization profile pending.
- Size
- 110B
- Memory
- 220GB
- Min Devices
- 4x
- Status
- Review required
THOX-ai/thox-cluster-100bCluster 200B
Meta · Llama 3.3
Exceeds standard MagStack memory; awaiting 8x cluster validation.
- Size
- 405B
- Memory
- 810GB
- Min Devices
- 8x
- Status
- Review required
THOX-ai/thox-cluster-200bGPT-OSS 120B
OpenAI · GPT-OSS
Memory footprint exceeds 8x MagStack budget; cloud-only until quantization or 16x stack lands.
- Size
- 120B
- Memory
- 240GB
- Min Devices
- 12x
- Status
- Cloud-only — under evaluation
gpt-oss:120bMixtral 8x22B
Mistral · Mixtral
Promising community reports; awaiting first-party Jetson AI Lab benchmark before approval.
- Size
- 141B
- Memory
- 90GB
- Min Devices
- 4x
- Status
- Review required
mixtral:8x22bCurrent device and runtime map
ThoxKey + ThoxBeam
ThoxKey is the portable runtime carrier. ThoxBeam serves signed assets from the key and uses the attached host GPU through WebGPU when supported. Large-model capability belongs to the host runtime, not the USB device itself.
ThoxAir
Compact wireless companion for role models, routing, orchestration, and delegated local-first work. Host or mesh execution is used when a workload exceeds the device budget.
ThoxMini
Compact local-first device for small role models and task-specific inference. Larger published models are companion or host workloads unless a device-specific benchmark says otherwise.
ThoxClip
Magnetic charging companion with onboard coordination compute. It can run small device roles and delegate heavier inference to authorized THOX hosts or MeshStack resources.
Host runtime and MeshStack
Desktop-class runtimes handle larger Ollama, llama.cpp, WebGPU, or accelerator-backed models. Multi-node jobs belong in the separate Cluster Models guide so a distributed requirement is never presented as single-device compatibility.
Quick Start Guide
# Pull a model from Ollama
ollama pull ministral-3:8b
# Run the model
ollama run ministral-3:8b
# For vision tasks, attach an image
ollama run ministral-3:8b "Analyze this medical image" /path/to/image.jpg
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.