Edge0-145B: Frontier AI Capabilities on Your Own Computer

Edge0-145B runs on one computer because only a small part of it has to sit on the graphics card at any moment. Edge0, a company that makes AI models for phones, headphones and PCs, announced the model on November 12 as frontier-level coding and agent capability on a personal computer. Its repository says the model is derived from Qwen3.8-Flash-Next, the open-weight model Qwen released in August (Qwen's launch post describes that base as 125 billion parameters plus a 51-billion-parameter n-gram embedding table). Edge0-145B is a mixture-of-experts model: each of its 48 layers holds 512 small expert networks, and a router sends each token to 10 of them plus one shared expert, so about 10 billion parameters do work per token. Edge0 compresses each expert's input to a narrower width, which turns every expert into a small fixed-size block of weights. According to the model card, the engine keeps a few thousand of those blocks on the GPU and streams the rest from system RAM, and the 51.2-billion-parameter n-gram table, which the model looks up rather than computes with, is memory-mapped from the SSD. A small draft model also proposes several tokens at a time for the main model to check, so each read of the weights yields more output.

The hardware the release files document is narrower than the blog's laptop pitch. The blog cites 29GB peak memory and more than 80 tokens per second on a compatible laptop such as a MacBook Pro, calling the 29GB a measured internal peak without saying which memory it counts, and says day-one support covers macOS, Windows and Linux, including Apple M-series chips. The model page on GitHub lists the requirements as Linux or Windows with an NVIDIA GPU and CUDA 13. Its speed table covers an RTX 5090 D, an RTX 4090, an RTX Spark N1X and an Apple M2 Ultra, none of them a MacBook, and it says the M2 Ultra figures do not establish macOS availability. On the RTX 5090 D, which has 24GB of video memory, the NVFP4 build generated 79.6 tokens per second in chat and 98.3 in an agent workload, and the page says that card holds the NVFP4 build as 23.25 GiB on the GPU plus 38.27 GiB in system RAM. With the same card capped at 16 GiB of video memory and 20 GiB of RAM, the smaller Q3_K build still produced 36.8 tokens per second in chat, which Edge0 labels a test budget rather than a minimum requirement. What the files document is Linux or Windows on a single high-end NVIDIA GPU, with much of the model kept in system RAM rather than on the card.

The quality claims are Edge0's own. Its table puts Edge0-145B ahead of its Qwen base on two agent and coding tests (64.6 against 58.7 on DeepSWE 1.1, 86.5 against 83.3 on Terminal-Bench 2.1) and slightly behind it on the GPQA-Diamond reasoning test (90.4 against 91.7). The blog says the Edge0, Qwen, DeepSeek and Claude scores come from Edge0's internal evaluations, and that the setup and reproduction instructions are being released with a technical report; the repository's README links only a September report on the earlier 35B model. The blog and the GitHub page also disagree on the smaller Edge0-120B, listing 53.6 and 51.3 on DeepSWE, so treat the scores as provisional until someone else runs them. The open-source label fits the inference framework, which is Apache-2.0 licensed. The weights come under the Edge0 Community Model License together with Qwen's license: personal, research and internal use are allowed, but selling or commercially bundling the model, running a commercial model service or offering a commercial coding assistant on it needs a separate written license.