Edge0 AI has open-sourced Edge0-145B, a 145-billion-parameter model the company calls the world’s first open-source on-device AI model at that size. The weights landed on Hugging Face on 11 November, the GitHub release followed the same day, and the official blog post arrived on 12 November. Alongside it came Edge0-120B and the company’s unified inference framework and deployment tools.
The architecture is what makes the size practical: Edge0-145B is a latent sparse mixture-of-experts model with 145 billion total parameters but only 10.12 billion active per token, across 48 layers with 512 experts per layer. It has a 256K-token context window and was trained natively in 4-bit. The released builds are compact for the class — the Q3_K build weighs 40.81 GiB and the NVFP4 build 47.79 GiB — and per the model card, on a single RTX 5090 the Q3_K build decodes at 121.2 tokens per second, while NVFP4 runs at 98.3 tokens per second.
On laptops the numbers are striking too: per the company’s blog, a MacBook Pro runs the model with 29GB peak memory, over 1,100 tokens per second of prefill and over 80 tokens per second of decode, at 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 and 5090 GPUs. The model-page benchmarks — terminal-bench 2.1 at 86.5 versus 78.2 for Claude Opus 4.6 (Max), GPQA-Diamond at 90.4 and deepswe 1.1 at 64.6 — are vendor-reported, so treat them as claims until independent tests reproduce them.
Adoption is the early tell: by 14 November the two Edge0-145B builds had 61,192 combined downloads on Hugging Face and the GitHub repo had 11,379 stars. The pitch is straightforward — no cloud token or API fees, data stays on your own computer, and core capabilities keep working offline — though hardware and electricity still cost money. For scale, the earlier Edge0-35B targeted phones under 2.5GB, while before September 2026 most on-device models stayed below 10 billion parameters. Open weights also mean you can pin versions, audit model files and run on infrastructure you control.