Edge0-145B drops local MoE with caveats

Edge0 proves that sparse MoE weight-streaming can push 145-billion-parameter models onto desktop GPUs, but the pitch relies on optimistic memory slicing and unverified internal benchmarks.

• –Tiered memory architecture: While only ~10B parameters activate per token and marketing highlights 29GB peak memory, running the model requires 23.25 GiB VRAM plus 38.27 GiB host RAM on an RTX 5090, alongside a 48GB SSD vocabulary table.
• –Hardware and platform reality: Advertised as compatible with laptops including MacBooks, the initial NVFP4 release is Blackwell CUDA-only with a Windows-only desktop client and zero current macOS runtime.
• –Base model and unverified evals: Built on Qwen3.8-Flash-Next with the edge0-strata engine, benchmark scores (64.6 DeepSWE, 90.4 GPQA-Diamond) remain vendor-reported without reproducible technical reports.
• –Licensing restrictions: Framed as open source, Edge0's weights require a proprietary Edge0 Community Model License on top of Qwen terms, requiring commercial waivers for model services or coding tools.
• –Developer bottom line: If you have a 24GB NVIDIA card and 64GB+ RAM, the weight-streaming architecture offers zero-token-cost local agent loops; Mac users should hold off entirely.