ORIX · On-device models
Models that run on your own device.
The open models ORIX Astra can download and run locally, on a Mac, an iPhone or an iPad. Each one was really downloaded, loaded and run before it was listed, and the figures below are the ones recorded that day.
44 models · 33 served by this site · 168 GB of weights · catalogue of November 22, 2026
Larger models, for the Mac 11
From 12 to 32 billion parameters. They need the memory of a Mac and are offered by ORIX Astra for the Mac only.
- Mac
Mistral Nemo 12B Instruct
Mistral AI · 4-bit · apache-2.0
- Parameters
- 12.2 B
- Download
- 6.4 GB
- Mac memory
- 16 GB or more
- Measured speed
- 75.1 tok/s
Text1,024,000 tokens of context
Served by www.orixlab.dev · mlx-community/Mistral-Nemo-Instruct-2407-4bit
- Mac
Phi-4
Microsoft · 4-bit · mit
- Parameters
- 14.7 B
- Download
- 7.7 GB
- Mac memory
- 16 GB or more
- Measured speed
- 61.6 tok/s
Text16,384 tokens of context
Served by www.orixlab.dev · mlx-community/phi-4-4bit
- Mac
DeepSeek R1 Distill 14B
DeepSeek · 4-bit · mit
- Parameters
- 14.8 B
- Download
- 7.7 GB
- Mac memory
- 16 GB or more
- Measured speed
- 57.5 tok/s
TextReasoning131,072 tokens of context
Served by www.orixlab.dev · mlx-community/DeepSeek-R1-Distill-Qwen-14B-4bit
- Mac
Qwen2.5 Coder 14B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 14.8 B
- Download
- 7.7 GB
- Mac memory
- 16 GB or more
- Measured speed
- 61.3 tok/s
TextTool use32,768 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen2.5-Coder-14B-Instruct-4bit
- Mac
Qwen3 14B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 14.8 B
- Download
- 7.8 GB
- Mac memory
- 16 GB or more
- Measured speed
- 57.4 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-14B-4bit
- Mac
Mistral Small 3.2 24B
Mistral AI · 4-bit · apache-2.0
- Parameters
- 24.0 B
- Download
- 12.4 GB
- Mac memory
- 24 GB or more
- Measured speed
- 38.0 tok/s
Text131,072 tokens of context
Served by www.orixlab.dev · mlx-community/Mistral-Small-3.2-24B-Instruct-2506-4bit
- Mac
Qwen3 30B A3B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 30.5 B
- Download
- 16.0 GB
- Mac memory
- 32 GB or more
- Measured speed
- 135.6 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-30B-A3B-4bit
- Mac
Qwen3 Coder 30B A3B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 30.5 B
- Download
- 16.0 GB
- Mac memory
- 32 GB or more
- Measured speed
- 94.1 tok/s
TextTool use262,144 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit
- Mac
DeepSeek R1 Distill 32B
DeepSeek · 4-bit · mit
- Parameters
- 32.8 B
- Download
- 17.2 GB
- Mac memory
- 32 GB or more
- Measured speed
- 27.8 tok/s
TextReasoning131,072 tokens of context
Served by www.orixlab.dev · mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit
- Mac
Qwen2.5 Coder 32B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 32.8 B
- Download
- 17.2 GB
- Mac memory
- 32 GB or more
- Measured speed
- 27.4 tok/s
TextTool use32,768 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen2.5-Coder-32B-Instruct-4bit
- Mac
Qwen3 32B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 32.8 B
- Download
- 17.2 GB
- Mac memory
- 32 GB or more
- Measured speed
- 27.8 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-32B-4bit
For Mac, iPhone and iPad 22
Compact models offered on every device. On an iPhone, the app checks the device’s memory before it suggests one.
- MaciPhone · iPad
SmolLM2 135M Instruct
Hugging Face (HuggingFaceTB) · 8-bit · apache-2.0
- Parameters
- 135 M
- Download
- 141 MB
- Mac memory
- 8 GB or more
- Measured speed
- 597.4 tok/s
Text8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolLM2-135M-Instruct-8bit
- MaciPhone · iPad
SmolLM2 135M Instruct
Hugging Face (HuggingFaceTB) · 16-bit floats · apache-2.0
- Parameters
- 135 M
- Download
- 261 MB
- Mac memory
- 8 GB or more
- Measured speed
- 553.0 tok/s
Text8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolLM2-135M-Instruct
- MaciPhone · iPad
Qwen2.5 0.5B Instruct
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 494 M
- Download
- 276 MB
- Mac memory
- 8 GB or more
- Measured speed
- 607.6 tok/s
Text32,768 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen2.5-0.5B-Instruct-4bit
- MaciPhone · iPad
SmolLM2 360M Instruct
Hugging Face (HuggingFaceTB) · 6-bit · apache-2.0
- Parameters
- 362 M
- Download
- 285 MB
- Mac memory
- 8 GB or more
- Measured speed
- 524.3 tok/s
Text8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolLM2-360M-Instruct-6bit
- MaciPhone · iPad
Qwen3 0.6B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 600 M
- Download
- 335 MB
- Mac memory
- 8 GB or more
- Measured speed
- 479.4 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-0.6B-4bit
- MaciPhone · iPad
SmolVLM2 256M
Hugging Face (HuggingFaceTB) · 16-bit floats · apache-2.0
- Parameters
- 256 M
- Download
- 494 MB
- Mac memory
- 8 GB or more
- Measured speed
- 584.0 tok/s
TextImages8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolVLM2-256M-Video-Instruct-mlx
- MaciPhone · iPad
Qwen3 0.6B (8-bit)
Alibaba Cloud (Qwen) · 8-bit · apache-2.0
- Parameters
- 600 M
- Download
- 619 MB
- Mac memory
- 8 GB or more
- Measured speed
- 436.6 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-0.6B-8bit
- MaciPhone · iPad
SmolLM2 360M Instruct
Hugging Face (HuggingFaceTB) · 16-bit floats · apache-2.0
- Parameters
- 362 M
- Download
- 695 MB
- Mac memory
- 8 GB or more
- Measured speed
- 424.2 tok/s
Text8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolLM2-360M-Instruct
- MaciPhone · iPad
Qwen2.5 Coder 1.5B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 1.5 B
- Download
- 839 MB
- Mac memory
- 8 GB or more
- Measured speed
- 369.4 tok/s
Text32,768 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit
- MaciPhone · iPad
Qwen3 1.7B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 1.7 B
- Download
- 938 MB
- Mac memory
- 8 GB or more
- Measured speed
- 350.3 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-1.7B-4bit
- MaciPhone · iPad
Qwen3 1.7B (4-bit DWQ)
Alibaba Cloud (Qwen) · 4-bit (DWQ) · apache-2.0
- Parameters
- 1.7 B
- Download
- 938 MB
- Mac memory
- 8 GB or more
- Measured speed
- 327.6 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-1.7B-4bit-DWQ-053125
- MaciPhone · iPad
DeepSeek R1 Distill 1.5B
DeepSeek · 4-bit · mit
- Parameters
- 1.8 B
- Download
- 965 MB
- Mac memory
- 8 GB or more
- Measured speed
- 340.5 tok/s
TextReasoning131,072 tokens of context
Served by www.orixlab.dev · mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-4bit
- MaciPhone · iPad
SmolVLM2 500M
Hugging Face (HuggingFaceTB) · 16-bit floats · apache-2.0
- Parameters
- 507 M
- Download
- 973 MB
- Mac memory
- 8 GB or more
- Measured speed
- 417.3 tok/s
TextImages8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolVLM2-500M-Video-Instruct-mlx
- MaciPhone · iPad
SmolLM3 3B
Hugging Face (HuggingFaceTB) · 4-bit · apache-2.0
- Parameters
- 3.1 B
- Download
- 1.6 GB
- Mac memory
- 8 GB or more
- Measured speed
- 212.7 tok/s
TextReasoning65,536 tokens of context
Served by www.orixlab.dev · mlx-community/SmolLM3-3B-4bit
- MaciPhone · iPad
Phi-3.5 mini Instruct
Microsoft · 4-bit · mit
- Parameters
- 3.8 B
- Download
- 2.0 GB
- Mac memory
- 8 GB or more
- Measured speed
- 187.5 tok/s
Text131,072 tokens of context
Served by www.orixlab.dev · mlx-community/Phi-3.5-mini-instruct-4bit
- MaciPhone · iPad
Phi-4 mini Instruct
Microsoft · 4-bit · mit
- Parameters
- 3.8 B
- Download
- 2.0 GB
- Mac memory
- 8 GB or more
- Measured speed
- 190.6 tok/s
Text131,072 tokens of context
Served by www.orixlab.dev · mlx-community/Phi-4-mini-instruct-4bit
- MaciPhone · iPad
Qwen3 4B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 4.0 B
- Download
- 2.1 GB
- Mac memory
- 8 GB or more
- Measured speed
- 176.2 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-4B-4bit
- MaciPhone · iPad
SmolLM2 1.7B Instruct
Hugging Face (HuggingFaceTB) · 16-bit floats · apache-2.0
- Parameters
- 1.7 B
- Download
- 3.2 GB
- Mac memory
- 8 GB or more
- Measured speed
- 138.0 tok/s
Text8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolLM2-1.7B-Instruct
- MaciPhone · iPad
DeepSeek R1 Distill 7B
DeepSeek · 4-bit · mit
- Parameters
- 7.6 B
- Download
- 4.0 GB
- Mac memory
- 16 GB or more
- Measured speed
- 108.5 tok/s
TextReasoning131,072 tokens of context
Served by www.orixlab.dev · mlx-community/DeepSeek-R1-Distill-Qwen-7B-4bit
- MaciPhone · iPad
Qwen2.5 Coder 7B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 7.6 B
- Download
- 4.0 GB
- Mac memory
- 16 GB or more
- Measured speed
- 115.9 tok/s
TextTool use32,768 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen2.5-Coder-7B-Instruct-4bit
- MaciPhone · iPad
SmolVLM2 2.2B
Hugging Face (HuggingFaceTB) · 16-bit floats · apache-2.0
- Parameters
- 2.2 B
- Download
- 4.2 GB
- Mac memory
- 8 GB or more
- Measured speed
- 130.4 tok/s
TextImages8,192 tokens of context
Served by www.orixlab.dev · mlx-community/SmolVLM2-2.2B-Instruct-mlx
- MaciPhone · iPad
Qwen3 8B
Alibaba Cloud (Qwen) · 4-bit · apache-2.0
- Parameters
- 8.2 B
- Download
- 4.3 GB
- Mac memory
- 16 GB or more
- Measured speed
- 103.5 tok/s
TextReasoningTool use40,960 tokens of context
Served by www.orixlab.dev · mlx-community/Qwen3-8B-4bit
iPhone and iPad only 11
Models whose licence does not let this site re-host them. The iPhone app downloads them from Hugging Face; they are not offered on the Mac, which downloads from this site only.
- iPhone · iPad
Gemma 3 270M
Google · 4-bit · gemma
- Parameters
- 268 M
- Download
- 181 MB
- Mac memory
- –
- Measured speed
- 618.7 tok/s
Text32,768 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/gemma-3-270m-it-4bit
- iPhone · iPad
LFM2 1.2B
Liquid AI · 4-bit · lfm1.0
- Parameters
- 1.2 B
- Download
- 633 MB
- Mac memory
- –
- Measured speed
- 567.0 tok/s
Text128,000 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/LFM2-1.2B-4bit
- iPhone · iPad
LFM2.5 1.2B Instruct
Liquid AI · 4-bit · lfm1.0
- Parameters
- 1.2 B
- Download
- 633 MB
- Mac memory
- –
- Measured speed
- 559.2 tok/s
TextTool use128,000 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/LFM2.5-1.2B-Instruct-4bit
- iPhone · iPad
Llama 3.2 1B Instruct
Meta · 4-bit · llama3.2
- Parameters
- 1.2 B
- Download
- 680 MB
- Mac memory
- –
- Measured speed
- 545.7 tok/s
TextTool use131,072 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/Llama-3.2-1B-Instruct-4bit
- iPhone · iPad
Gemma 3 1B (QAT)
Google · 4-bit (QAT) · gemma
- Parameters
- 1.0 B
- Download
- 736 MB
- Mac memory
- –
- Measured speed
- 325.1 tok/s
Text32,768 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/gemma-3-1b-it-qat-4bit
- iPhone · iPad
LFM2 2.6B
Liquid AI · 4-bit · lfm1.0
- Parameters
- 2.6 B
- Download
- 1.4 GB
- Mac memory
- –
- Measured speed
- 270.8 tok/s
TextTool use128,000 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/LFM2-2.6B-4bit
- iPhone · iPad
Qwen2.5 3B Instruct
Alibaba Cloud (Qwen) · 4-bit · qwen-research
- Parameters
- 3.1 B
- Download
- 1.6 GB
- Mac memory
- –
- Measured speed
- 219.2 tok/s
TextTool use32,768 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/Qwen2.5-3B-Instruct-4bit
- iPhone · iPad
Qwen2.5 Coder 3B
Alibaba Cloud (Qwen) · 4-bit · qwen-research
- Parameters
- 3.1 B
- Download
- 1.6 GB
- Mac memory
- –
- Measured speed
- 219.1 tok/s
TextTool use32,768 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/Qwen2.5-Coder-3B-Instruct-4bit
- iPhone · iPad
Llama 3.2 3B Instruct
Meta · 4-bit · llama3.2
- Parameters
- 3.2 B
- Download
- 1.7 GB
- Mac memory
- –
- Measured speed
- 229.4 tok/s
TextTool use131,072 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/Llama-3.2-3B-Instruct-4bit
- iPhone · iPad
Qwen2.5-VL 3B
Alibaba Cloud (Qwen) · 4-bit · qwen-research
- Parameters
- 3.8 B
- Download
- 2.9 GB
- Mac memory
- –
- Measured speed
- 219.8 tok/s
TextImages128,000 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/Qwen2.5-VL-3B-Instruct-4bit
- iPhone · iPad
Gemma 3 4B
Google · 4-bit · gemma
- Parameters
- 4.3 B
- Download
- 3.2 GB
- Mac memory
- –
- Measured speed
- 120.8 tok/s
TextImages2,048 tokens of context
Downloaded from Hugging Face by the iPhone app · mlx-community/gemma-3-4b-it-4bit
- 01
One source on the Mac
ORIX Astra for the Mac downloads every model from this site and from nowhere else. Each file is checked against its published SHA-256 before a model can be loaded.
- 02
Open licences
Only models under Apache-2.0 or MIT are hosted here. A model keeps the licence of its publisher; the name of its original repository is shown on each card.
- 03
Private by design
A downloaded model runs on your device. With an on-device model, what you write and what it answers never leave it.
Speeds were measured on one machine, Mac17,6 (Apple M5 Max, 48 GB). Another chip, or less memory, gives other numbers. “Mac memory” is an estimate from the size of the weights, not a measurement on a Mac of that size.