145B Model Runs on a Laptop

🔥 What happened A team open-sourced Edge0-145B and Edge0-120B, calling it the first on-device model with 145 billion parameters. It reportedly runs on a laptop at 29GB peak memory and over 80 tokens/s. 💡 Why it matters On-device models were stuck under 10B parameters; Gemini Nano sits at 1.8B. Edge0 shares large matrices across 64 experts and activates only ~10B parameters per step — that's the real trick. ⚡ Our take Impressive, but the benchmarks are self-reported. Anyone with 29GB RAM and 120W on a laptop isn't running a normal machine anyway.