Edge0 has released Edge0-145B and a smaller Edge0-120B as open models designed to run on personal computers, according to a company blog post published November 12. The release includes an inference framework, deployment tools and a desktop app that the company says runs offline with no account or usage quota.
Edge0-145B is a sparse mixture-of-experts model with 145 billion total parameters and about 10 billion active per inference step. Edge0 says experts are grouped in sets of 64 that share two large matrices, and the model was trained in its on-device form from the start rather than compressed afterwards. It also uses speculative decoding. The company reports peak memory use of 29GB, prefill above 1,100 tokens per second, decoding above 80 tokens per second, about 120W of typical power draw, and a 128K to 256K token context. To keep memory low, about 40GB of expert weights sit in system memory and about 48GB of vocabulary data sits on the SSD.
Day-one support covers macOS, Windows and Linux, including Apple M-series chips and NVIDIA's RTX 4090, RTX 5090 and RTX Spark hardware. Edge0 cites local use on a 48GB laptop and says very low-compute devices cannot run the model.
Edge0 reports scores of 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. Its six-benchmark average of 72.8 compares with 71.7 for Qwen3.8-Flash-Next, which Edge0 estimates needs about 70GB at 4-bit. The company says most of these scores come from internal evaluations, that baseline memory figures are estimates, and that a technical report with the full evaluation setup is still to come.
Weights are on Hugging Face as NVFP4 checkpoints. The blog post calls the models open source but does not name a license, so developers should check the terms on the model pages before building on them. If the numbers hold up under independent testing, the release would put a model of this size within reach of a high-end laptop without a cloud API.