Edge0-145B is available to download. Here’s what a PC needs to run it

Edge0 announced two downloadable coding models on November 12, with weights available on Hugging Face. The released setup for Edge0-145B and Edge0-120B uses one NVIDIA GPU on a Windows or Linux PC with CUDA 13. Developers with a compatible system can load either model and send code-generation requests from their tools to a server on the same computer.

The announcement also listed Apple's M-series Macs among the systems supported at launch. The repository's November 13 release entry and its compatibility table list Windows and Linux for both new models. The model card includes a test on an Apple M2 Ultra and says that measurement does not establish macOS availability. Edge0's Mac guide describes how to load the earlier 8B and 35B models.

The GPU, system memory and SSD do different jobs

Edge0-145B gets its name from its 145 billion parameters. The model is derived from Qwen3.8-Flash-Next, and Edge0 says about 10.12 billion parameters are active when it processes a token. Different parts of the model handle different inputs and share some of their larger matrices. The company says it keeps frequently used parts in GPU memory and transfers others from system memory as needed.

The model's Q3_K weights are listed at 40.81 GiB in the technical documentation. The full folder on Hugging Face is about 105 GB, including a 51.2 GB table that the model consults during processing. The runtime reads the relevant parts of that table from the SSD by default. The weights figure and the folder size describe different things, so the full download needs more storage than the weights alone suggest.

Edge0's run guide specifies at least 16 GiB of GPU memory and 20 GiB of system memory. The model card's performance row with those numbers came from a 24 GB RTX 5090 D whose memory budget was capped for the test. Edge0 labels it a test condition and gives a separate peak figure of 29 GB in its launch article for a particular laptop setup.

Two models, different results on the same card

Edge0 reports that the 145B Q3_K build generated 99.7 tokens per second in chat and 121.2 in agent tasks on an RTX 5090 D 24 GB. With the memory cap in place, those figures were 36.8 and 44.9. These are rates for generating output. The company measures input processing separately, using fixed prompt lengths.

The 120B Q3_K weights take up 27.98 GiB. With the same memory limits, Edge0 recorded 62.4 tokens per second for that model in chat. The larger model scored higher on the company's DeepSWE 1.1 software engineering evaluation.

Edge0's software can open an interactive chat or serve requests on the same computer. The Windows desktop app can load a model from a downloaded folder, according to the project documentation. Developers can also connect clients to the local server. The model generates the responses, and a connected tool handles actions such as applying a change to a file.

The weights carry their own terms

Edge0 publishes the framework code under Apache 2.0. Its Community Model License covers the contributions it made to the weights. It permits personal use, research, internal use and commercial use of generated outputs. The license reserves paid model distribution, commercial inference services and independent commercial coding or office assistants for a separate written agreement.

The repository also carries the Qwen license for the upstream material. Edge0's license says permission from Edge0 leaves any applicable Qwen obligations in place.