New Large Language Model Released for On-Device Use

Edge0 has announced the release of Edge0-145B, a new open-source large language model designed to bring frontier AI capabilities to local hardware. The model is part of a “latent-MoE” (Mixture of Experts) series that includes a 120B parameter version. According to technical details provided by the developer, the model is derived from Qwen3.8-Flash-Next and was trained natively at 4 bits.

The release aims to facilitate high-performance AI execution on individual computers rather than relying on cloud-based infrastructure. Documentation for the project has been made available via GitHub at the Edge0-AI repository.

Performance and Benchmarks

Initial discussions on community platforms like Reddit have highlighted various performance metrics. Some users have pointed to measured benchmarks that suggest the model performs well across standard testing suites. Additionally, speed benchmarks have been published on Hugging Face, detailing execution speeds across different hardware configurations, including specific data for 24 GB VRAM setups.

While some community members have expressed optimism regarding the model’s potential to compete with existing large-scale models, these assessments remain unverified by independent third-party testing.

Technical Scrutiny and Community Reception

The release has met with a mixed reception within developer communities. While some users have characterized the model as a significant development for the open-source space, others have raised questions regarding the nature of its training. Specifically, some commentators on Reddit have questioned whether the model is a novel architecture trained from scratch or a fine-tuned version of existing technology, such as a QFN fine-tune.

There has also been some skepticism regarding the identity and legitimacy of the Edge0 group, with some users noting a lack of prior public presence for the developer. As with many new releases in the rapidly evolving open-source AI sector, the technical claims and the origin of the training data have not yet been independently confirmed by industry analysts.