Edge0 released Edge0-145B on November 12, an open-weight model with 145 billion total parameters that the company says can run on a personal computer. It also released a smaller Edge0-120B, along with an inference framework, deployment tools and a desktop app.
The model uses a mixture-of-experts design in which roughly 10 billion parameters are active on each inference step. Edge0 reports a peak memory footprint of about 29GB, achieved by keeping an expert library of about 40GB in system memory and around 48GB of vocabulary data on SSD. The company cites a 48GB laptop such as a MacBook Pro as sufficient, with day-one support for macOS, Windows and Linux on Apple M-series chips and Nvidia RTX Spark, RTX 4090 and RTX 5090 hardware. It reports decoding speeds above 80 tokens per second, prefill above 1,100 tokens per second, and a context window of 128K to 256K tokens.
On Edge0's own internal evaluations, the model scores 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 35.4 on HLE, 90.4 on GPQA-Diamond, 80.6 on IFBench and 79.3 on AA-LCR. These figures have not been independently reproduced.
Weights are published on Hugging Face as Edge0/Edge0-145B-NVFP4 and Edge0/Edge0-120B-NVFP4, with code and deployment guides at the Edge0-AI/Edge0 GitHub repository. The Hugging Face model card lists the Edge0 Community Model License 1.0 for Edge0's own contributions and notes that Qwen materials remain under the Qwen Community License 1.0, so developers should check both licences before commercial use.
For teams that need private or offline inference, the release is an attempt to bring a model of this parameter count onto a single high-memory laptop by keeping most of its weights out of active computation and off GPU memory.