Edge0-145B Brings 100B+ Parameter AI to Local Computers

Edge0 released Edge0-145B, an open-source AI model with 145 billion parameters designed to run on personal computers without cloud dependence. The model achieves 80+ tokens per second on laptops while consuming 28GB of peak memory. It performs competitively with cloud models on coding (scoring 64.6 on DeepSWE 1.1), reasoning (90.4 on GPQA-Diamond), and instruction following (80.6 on IFBench), often exceeding lighter alternatives.

The key innovation is architecture designed for local constraints rather than adapted from cloud models. The design uses sparse experts with shared large matrices, so only approximately 10B of the model’s 145B parameters compute per inference step. Supporting infrastructure includes an approximately 40GB expert library in system memory and approximately 48GB of vocabulary data stored on SSD. This approach delivers a 72.8 mean score across six benchmarks with 29GB peak memory—roughly 40% of the estimated memory Qwen3.8-Flash-Next requires (approximately 70GB) while achieving similar performance.

The breakthrough represents a shift in on-device AI scaling. Prior models remained constrained below 10B parameters due to memory limits, with Apple Intelligence at approximately 3B and Google’s Gemini Nano at 1.8B to 3.25B. Edge0-145B demonstrates that 100B+ parameter inference is feasible on personal hardware when optimization spans the full stack from architecture through training, memory management, and inference scheduling.