Edge0 open-sources 145B on-device model claiming 29GB laptop memory and 80+ tokens/s

Edge0 released open weights for Edge0-145B and Edge0-120B, along with a unified inference framework and deployment tools. The company describes Edge0-145B as the first open-source on-device AI model with 145 billion parameters. Its published laptop results report 29GB peak memory, 1,100-plus tokens per second prefill, 80-plus tokens per second decode, about 120 watts typical power, and 128K to 256K context. The release also includes an Edge0 app for local use.

The source presents these as vendor-reported measurements on compatible hardware, such as a MacBook Pro, and says the model can sustain long tasks at essentially the same speed. It also says coding, reasoning, and instruction-following results are competitive with specific frontier cloud models, though the provided excerpt does not list benchmark tables. An earlier Edge0-35B phone result reports peak memory below 2.5GB and 37-plus tokens per second on an iPhone 18 Pro Max.

The company frames the release against a memory wall that kept many local models small: Apple Intelligence used about 3B parameters, Gemini Nano 1.8B and 3.25B, and a 7B model could need about 4GB. Edge0 argues that cloud workflows add token or API fees, sensitive-data restrictions, and network dependence. Its stated goal is to let developers reproduce results, evaluate or fine-tune the models, and integrate them into products without those recurring cloud constraints.

The vendor-reported numbers matter because they describe high-parameter coding, reasoning, and agent workflows running on user-controlled machines. Developers, researchers, and enterprises handling sensitive code or data could test local alternatives to hosted models. The open stack also shifts some burden to hardware: laptops need enough memory, power headroom, and storage. The source does not include third-party benchmark tables, so independent reproduction remains the next evidence step.