📅 November 14, 2026 • Edge0 AI
Edge0 AI has open-sourced Edge0-145B and Edge0-120B together with the company's unified inference framework and deployment tools, calling Edge0-145B the world's first open-source on-device AI model with 145 billion parameters (per Edge0 AI). Weights landed on Hugging Face on November 11, 2026 (05:22 UTC), the GitHub release v0.2.0 followed the same day (15:13 UTC), and the official blog post arrived November 12. Attention is real: by November 14 the two builds of Edge0-145B had 61,192 combined downloads on Hugging Face and the GitHub repo had 11,379 stars.
Architecture built for one GPU
Edge0-145B is a latent sparse MoE: 145B total parameters with only 10.12B active per token, 48 layers, 512 experts per layer with 10 routed plus 1 shared expert active, a 256K context window and native 4-bit training (per the GitHub docs and the HF model card). The released builds are compact for the class — Q3_K weighs 40.81 GiB, NVFP4 47.79 GiB. Per the model card on an RTX 5090, Q3_K decodes at 121.2 tok/s with 3554.9 tok/s 16K prefill, NVFP4 runs 98.3 tok/s decode and 3165.7 tok/s prefill. Sibling Edge0-120B: 120B total, 8.64B active.
Laptop numbers, vendor-reported
Per the company's blog, on a MacBook Pro the model runs with 29GB peak memory, 1100+ tok/s prefill, 80+ tok/s decode and about 120W typical power, with 128K–256K context at essentially the same speed. Day-one support covers macOS, Windows and Linux on Apple M-series laptops and RTX Spark/4090/5090 GPUs. Model-page benchmarks: terminal-bench 2.1 = 86.5 versus 78.2 for Claude Opus 4.6 (Max), GPQA-Diamond = 90.4, deepswe 1.1 = 64.6 — vendor-reported, so treat them as claims until you reproduce them on your own tasks.
What it changes for local AI
The pitch stays neutral: no cloud-model token or API fees, data can stay on the computer, core capabilities keep working offline — hardware and electricity still cost money (per the blog). For scale: the earlier Edge0-35B targeted phones below 2.5GB with 37+ tok/s on an iPhone 18 Pro Max (per the blog), and before September 2026 on-device models mostly stayed below 10B (Apple around 3B, Gemini Nano 1.8B and 3.25B per the blog). Open weights also mean you can pin versions, audit model files and run on infrastructure you control.
IMPACT
For teams building automation on their own hardware, the line between cloud-only and local model classes shifts: a 145B-parameter model under 48 GiB that decodes past 98 tok/s on one RTX 5090 makes private, offline agents plausible for real workloads — if the vendor numbers hold on your machine. Practical step: pull the NVFP4 or Q3_K build, run your actual workflows and compare against your current stack on quality, latency and electricity cost, not on marketing benchmarks alone. Adoption is the tell — the download and star counts above show where builders are already testing.
Sources: edge0.ai — huggingface.co — github.com — activepieces.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
