Local AI is the Future of
Everyday Productivity.
DynaMoE brings frontier-grade Mixture-of-Experts and dense AI models directly to your Mac. By dynamically streaming active expert weights from NVMe SSD on demand, you get fast, private, and unconstrained local intelligence without expensive hardware requirements.
macOS 15.0+ Free & Open Source on Apple Silicon
Built for unconstrained local intelligence.
From dynamic SSD memory-mapping to custom Metal Shading Language kernels, DynaMoE optimizes every layer of the inference stack.
Dynamic SSD Weight Streaming
DynaMoE memory-maps multi-gigabyte SafeTensors checkpoints and pages active expert weights directly from high-speed NVMe storage on demand—running 35B+ parameter models on consumer Macs without preloading all weights into RAM.
Automatic Dense & MoE Dual-Engine
Seamlessly runs both dense LLMs (e.g., Nanbeige 4.2 3B, Gemma, Ornith 1.5 9B, Qwen) and sparse Mixture-of-Experts architectures (e.g., Ornith 1.5 35B, GLM 5.3 Flash) with dynamic topology auto-detection.
Zero-Copy Metal GPU Pipeline
Engineered natively with custom Metal Shading Language (MSL) compute kernels (Q4/Q8 GEMV, FP8, SwiGLU, GQA) utilizing Apple Silicon's Unified Memory architecture for zero-copy GPU execution and rapid token generation.
100% Private & Native macOS App
Complete offline independence. Prompts, reasoning traces, and local Hugging Face model cache remain strictly on your Mac inside an intuitive SwiftUI experience with collapsible reasoning stream accordions.