Edge-optimized Recurrent-Depth Transformer for ARM64/Jetson devices. Pure PyTorch, no Triton kernels, no CUDA dependencies.
- 1B and 3B variants designed for Jetson Orin (8GB VRAM)
- Pure PyTorch — no custom kernels, works on ARM64
- Memory estimation —
config.estimate_memory()before deployment - GQA attention — standard Grouped Query Attention (MLA optional)
- Adaptive compute — ACT halting with configurable threshold
- MoE sparse experts — 16 routed experts, top-K routing
from open_mythos_edge import OpenMythosEdge, mythos_1b_edge
config = mythos_1b_edge() # ~1.6 GB
model = OpenMythosEdge(config)