A company called Edge0 published the weights for a 145-billion-parameter AI model on November 12, 2026, and says it runs on one computer. The v0.2.0 release notes list what it takes: Windows, an NVIDIA graphics card with at least 16 GB of its own memory, 20 GB of system RAM, and NVIDIA's CUDA Toolkit 13, the software layer the model runs through. In practice that means a gaming PC or a workstation rather than a typical laptop. Edge0 says the result is a model this size running without the cloud, with no per-token fees. We verified the release artifacts, the download, and the model pages; we didn't run the model ourselves. We also read the announcement against the repository, and one headline claim didn't hold.
What Edge0 released
Edge0-145B is a mixture-of-experts model: instead of using all 145 billion parameters on every word, it routes each token through a small subset of specialist parameter groups. Only about 10 billion parameters are active per token, which is how the memory bill stays within reach of one machine. It handles up to 256,000 tokens of context and was trained to run at 4-bit precision from the start.
The weights are on Hugging Face in two weight formats: Q3_K, a compressed 4-bit-class format (about 41 GB), and NVFP4, an NVIDIA-oriented 4-bit format (about 48 GB). The inference framework, edge0-infer, is open source on GitHub under Apache 2.0, and the v0.2.0 release ships a Windows chat app. A smaller sibling, Edge0-120B, uses the same design.
On benchmarks in Edge0's announcement, the model scored 86.5 on Terminal-Bench 2.1 and 64.6 on DeepSWE 1.1, both agent-and-coding tests, ahead of the named models Edge0 compared it against. The company ran those evaluations itself, including its rivals' scores. The model is also derived from Qwen3.8-Flash-Next, and Edge0's own table shows it trailing that base model on some knowledge benchmarks while leading on the agent ones.
Why a beginner might care
If you're deciding what local AI might fit your hardware, Edge0's own catalog now spans both ends: the company says its September 2026 Edge0-35B model targets phones and runs in about 3 GB of memory, and this 145B release aims at desktops with large NVIDIA cards. Matching the model to the machine is the decision that matters.
If you do own a qualifying machine, the math changes. Local inference has no token bill, so a project that takes fifty rounds of writing code, running it, and fixing errors costs the same as one that takes five. Prompts and files can stay on your drive, and a fully local setup can keep working without an internet connection.
What we checked on launch day
The release is downloadable. The v0.2.0 release page lists Windows x64 and Arm64 app packages. The Windows x64 download link's headers reported a file of about 264 MB. We did not install or open the app.
The weights are live. The Hugging Face pages for Edge0-145B-q3k and Edge0-145B-NVFP4 load, and the model card lists the weight sizes (40.81 GiB and 47.79 GiB), the architecture (48 layers, 512 experts, 10 routed plus 1 shared active per token), and the license.
The platform table contradicts one announcement claim. More on that below, because it's the honest catch.
We did not run the model. Our test machine is a Mac without an NVIDIA GPU, so the app-to-answer steps in our guide carry Edge0's documentation, labeled as untested.
The honest catch
Edge0's announcement describes the 145B model running on a compatible laptop, such as a MacBook Pro, with reported peak memory of 29 GB and decoding above 80 tokens per second. The company's own repository tells a narrower story: the platform table lists macOS with a dash for the 145B and 120B models, and the app release is Windows-only. The only Apple measurement in the repo is a Mac Studio class M2 Ultra at roughly 57 to 61 tokens per second, with a note that this doesn't establish macOS availability. We couldn't reproduce the MacBook claim or find a shipping artifact that supports it.
The 29 GB figure doesn't match the repo's own numbers either. The release notes put the practical floor at a 16 GB NVIDIA GPU plus 20 GB of system RAM, and the repo's constrained test shows the Q3_K build holding about 23 GB of video memory and 31 GB of system RAM. A laptop whose 8 or 16 GB of memory is shared between processor and graphics doesn't meet that bar, and that describes most laptops.
Two smaller catches: the benchmark table comes from Edge0's internal evaluations, so treat it as a starting point. And the weights aren't open source in the way software is. The Edge0 Community Model License permits personal use, research, and commercial use of what the model generates, but selling the model, running a paid model service, or building a commercial coding assistant on it takes a separate written license. Qwen's upstream terms apply too.
For a first local model tonight, Edge0-145B may be more than your machine can run. A smaller model matched to the hardware you already own is the practical start.
The full walkthrough
The guide carries the steps so this post can focus on what shipped and what we verified. It covers the hardware check first (what qualifies, what doesn't, and the 41 GB download before you run anything), then the download order, a starter agent job shaped for a local model, the license limits, and a dated record of what we did and didn't test: Edge0-145B: a big open model that skips the cloud.
FAQ
What is Edge0-145B?
It's an open-weights AI model published November 12, 2026, with 145 billion total parameters and about 10 billion active per token. It's built to run on a single computer with an NVIDIA graphics card, and it targets coding and agent-style work: planning a task, writing and running code, and fixing errors without a cloud service.
Can Edge0-145B run on a MacBook Pro?
Not at this release. Edge0's announcement mentions MacBook Pro-class laptops, but the company's own platform table lists macOS with a dash for the 145B and 120B models, and the v0.2.0 app is Windows-only. The repo's only Apple measurement is a Mac Studio class M2 Ultra, which the repo says doesn't establish macOS availability. Edge0's smaller 35B model lists macOS support in the repository.
Can you use Edge0-145B outputs commercially?
Yes. The Edge0 Community Model License permits commercial use of what the model generates. Building a product on the model itself is the different case: selling or bundling the model, running a paid model service, or offering a commercial coding or office assistant powered by it takes a separate written license, and Qwen's upstream terms also apply.
How much does Edge0-145B cost to use?
The weights are free to download and local inference has no per-token fees, so the cost is the hardware, the electricity, and the download time. The hardware is the expensive part: a qualifying NVIDIA GPU with 16 GB of its own memory typically means a gaming PC or workstation.
More on this: What Is an AI Agent? Explained Simply and What Is Context in AI? Why It Makes or Breaks Your Results.