Edge0-145B: Frontier AI Capabilities on Your Own Computer

Edge0 announces open-source releases of its 145-billion- and 120-billion-parameter models, along with an inference framework and deployment tools aimed at running advanced AI locally. The company reports that Edge0-145B uses 29GB of peak memory on compatible laptops, generates more than 80 tokens per second, processes prompts at over 1,100 tokens per second, and supports 128K-256K context at about 120W. Its approach combines shared matrices across sparse experts, a model trained specifically for local execution, and coordinated use of GPU memory, system memory, SSD storage, CPU, and GPU. About 10 billion of its 145 billion parameters are active per inference step; speculative decoding helps reduce the cost of loading model weights.

The model is presented as both a capable assistant and a local agent that can plan tasks, write and run code, debug failures, and inspect results visually. In company-reported evaluations, it scores 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, and 90.4 on GPQA-Diamond; its six-benchmark average is 72.8. Edge0 says this is competitive with named cloud and open models while using less memory - for example, compared with an estimated 70GB for Qwen3.8-Flash-Next. The release is framed as a way to reduce cloud fees, connectivity dependence, and exposure of sensitive data. The comparisons rely partly on internal evaluations, and some competing models’ memory figures are estimates rather than measurements.