Every model in the Magnitude catalog, tuned and tested for Apple Silicon, NVIDIA, AMD, and CPU.
23 models
Qwen ·
Large multimodal MoE preview of Qwen's next-generation architecture with strong agentic reasoning and low active compute.
69% Intelligence
Qwen ·
Large dense multimodal model with frontier-class agentic coding capability for its size.
59% Intelligence
NVIDIA ·
Workstation-class hybrid MoE vision model for reasoning, coding, and agentic workflows.
46% Intelligence
Google ·
Large dense Gemma model with the strongest measured coding score in its family.
33% Intelligence
OpenBMB ·
Multimodal MoE with a large knowledge footprint and low active compute for image, document, and video understanding.
33% Intelligence
NVIDIA ·
Compressed hybrid MoE vision model with 2-bit expert weights, sized for 24GB machines.
33% Intelligence
Qwen ·
Efficient MoE coding model with a large knowledge footprint and low active compute.
32% Intelligence
Meta ·
Dense multimodal model purpose-built for autonomous agentic tasks on consumer hardware.
30% Intelligence
Google ·
Mid-size MoE model balancing a substantial weight footprint with low active compute.
29% Intelligence
Liquid AI ·
Reasoning MoE for fast on-device agents, coding, and tool use.
27% Intelligence
Google ·
Mid-size dense model with native tool use, reasoning, and multimodal capability.
25% Intelligence
Liquid AI ·
Compact hybrid Liquid model for fast on-device agents and tool use, with a quantization-aware distilled (QAD) Q4.
24% Intelligence
Qwen ·
Compact dense model for machines where responsiveness and footprint matter most.
23% Intelligence
Liquid AI ·
Compact Liquid vision-language model for on-device image understanding and agents, with a quantization-aware distilled (QAD) Q4.
22% Intelligence
OpenBMB ·
Compact dense model with native long context, tool use, and reasoning for on-device work.
22% Intelligence
NVIDIA ·
Efficient hybrid MoE model for local reasoning, coding, and agentic workflows.
22% Intelligence
Qwen ·
Small dense model with a substantial capability gain over the 4B tier.
19% Intelligence
Google ·
Small dense model with per-layer embeddings that materially improves capability over the E2B tier while remaining practical on memory-constrained machines.
15% Intelligence
Liquid AI ·
Compact dense Liquid model tuned for fast on-device tool use and coding workflows.
15% Intelligence
OpenBMB ·
Compact dense model for on-device reasoning and tool use.
15% Intelligence
Google ·
Very small dense model with per-layer embeddings optimized for on-device use.
14% Intelligence
Liquid AI ·
Low-active-parameter Liquid MoE optimized for fast local reasoning and tool use.
13% Intelligence
OpenBMB ·
Compact multimodal model for on-device image, document, and video understanding.
10% Intelligence
No models match. Clear a filter to see more.
How a model compares with the most capable model available. It is the model’s Artificial Analysis Intelligence Index score, a composite of reasoning, coding, math, and knowledge benchmarks, as a percentage of the top model on the index. Where Artificial Analysis has not fully measured a model, Magnitude uses an estimate.
Quantization levels. Lower numbers mean smaller downloads and less memory at a small cost in fidelity. Variants marked QAT were trained with quantization in mind and keep more fidelity at the same size. Magnitude recommends the highest-fidelity variant that fits your memory with room for context.
Weights plus every companion file Magnitude installs: the vision projector for image-capable models and the draft model for speculative decoding. They are the exact files pinned by the catalog’s locked Hugging Face commits.
Magnitude profiles your hardware and estimates tokens per second for every model before you download. The sizes here are what you download; the app gives you the real answer for your chip and memory, including room for context.
Catalog data from magnitudedev/magnitude, updated Oct 28, 2026.