Edge0-145B: A 145B Local Model That Needs an NVIDIA GPU, Not a Mac
Edge0-145B is a 145 billion parameter sparse model from Edge0 that activates about 10 billion parameters per step and is built to run on a single personal machine. The company reports 80+ tokens per second and coding scores above several larger models, but today it needs Linux or Windows with an NVIDIA GPU, and commercial use needs a separate license.
On this page
What Edge0 released
Edge0 published Edge0-145B and a smaller Edge0-120B with an inference framework and a desktop app, under the headline of frontier AI capabilities on your own computer. The company, which previously released Edge0-35B in September, says it now has more than 5 million devices in production and over 15 million downloads of its open models.
Both new models are sparse: experts are grouped so that each group of 64 shares two large matrices while every expert keeps a small set of its own weights, and only about 10 billion of the 145 billion parameters work on each step. Edge0 says the models were trained in this on-device layout from the start rather than compressed after the fact, and that context runs to between 128K and 256K tokens. The project README adds an important detail the launch post plays down: both models are derived from Qwen3.8-Flash-Next.
The hardware story is narrower than the headline
The launch post lists a 29GB peak memory figure, 1,100+ tokens per second of prefill and 80+ tokens per second of decode, and says day-one support covers macOS, Windows, and Linux, including Apple M-series chips. The model card and README tell a more cautious story. The release path for the 145B and 120B models targets Linux or Windows with an NVIDIA GPU and CUDA 13, and the README lists no macOS support for them. Edge0 does publish Apple M2 Ultra measurements, but says they "do not establish macOS release availability".
The benchmarks on the model card were run on an RTX 5090 D with 24GB of video memory, where the NVFP4 build used 23.25 GiB of VRAM plus 38.27 GiB of system RAM. That split explains how a model whose weight files are 47.79 GiB (NVFP4) or 40.81 GiB (Q3_K) can report a smaller peak: frequently used experts sit on the GPU and the rest wait in system memory. By our reading of those figures, that means a high-end NVIDIA card plus generous system RAM, not just any laptop. Byteiota reached the same conclusion, pointing Apple Silicon owners to the 35B model for now.
Open weights, but not open for business
Edge0 calls the release open source, and the framework code is Apache 2.0. The weights are a different matter. Edge0's own contributions use the Edge0 Community Model License 1.0, and because the model comes from Qwen, the Qwen Community License 1.0 applies as well. Selling or bundling the model, running a commercial model service, or building an independent commercial coding or office assistant on it needs separate written permission from Edge0.
For personal use and private experiments that is fine. For a startup planning to ship it inside a product, it is a real constraint, and it puts Edge0-145B in a different category from Apache 2.0 releases.
How much to trust the scores
Edge0 reports 86.52 on Terminal-Bench 2.1, ahead of Qwen3.8-Flash-Next at 83.3 and Claude Opus 4.6 at 78.2, and 64.6 on DeepSWE 1.1. These are the company's own evaluations, and it says the full setup and reproduction instructions will arrive with a later technical report. Independent reproduction has not happened at scale yet, and commenters have asked whether the scores came from the quantized local build or from higher precision weights.
Beating its own base model would be a notable result for a 4-bit design, which is exactly why it needs outside testing before anyone treats it as settled.
What local AI users should do
If you already own an RTX 4090 or 5090 class card with plenty of system RAM, Edge0-145B is one of the most interesting large models you can try at home this week, and the engine exposes OpenAI and Anthropic compatible endpoints for existing tools. Mac owners should wait for a confirmed macOS build, and anyone with commercial plans should read both licenses first.