Edge0-145B is an open-weight AI model with 145 billion parameters that activates only about 10 billion per step, so it can run on a high-end laptop or a single consumer GPU. Edge0 released it, with a smaller Edge0-120B and a desktop app, on November 12, 2026.
Edge0-145B at a glance
Price | Free weights; commercial services need a separate license |
|---|---|
Size | 145B total, ~10B active per step |
Memory | ~29GB peak (Edge0's laptop figure) |
Speed | 1,100+ tokens/s prefill, 80+ tokens/s decode (Edge0's figures) |
Context | 128K-256K tokens |
Download | 47.79 GiB (NVFP4) or 40.81 GiB (Q3_K) |
License | Edge0 Community Model License 1.0 (plus Qwen Community License) |
What Edge0-145B is
It's a mixture-of-experts model: most of its parameters sit idle on any given step, which keeps memory and compute low. Edge0 pitches it as frontier-class AI for coding, agents and office work that runs offline with no account or usage quota. It's built on a Qwen base model, so Qwen's license terms apply alongside Edge0's.
Benchmarks (Edge0's own)
Edge0 reports 86.5 on Terminal-Bench 2.1, 64.6 on DeepSWE 1.1, 90.4 on GPQA-Diamond, 80.6 on IFBench, 79.3 on AA-LCR and 35.4 on HLE, a 72.8 average across the six. These come from Edge0's internal evaluations and a technical report is still to come, so treat them as claims until independent tests land.
Hardware you'll need
Edge0's announcement targets compatible laptops such as recent MacBook Pros and NVIDIA RTX 4090 and 5090 machines, with day-one support listed for macOS, Windows and Linux. The model card is more conservative: its reported setup is an NVIDIA RTX 5090 D with 24 GB of VRAM, using about 23 GiB of VRAM plus 38 GiB of system RAM, with a constrained 16 GiB VRAM / 20 GiB RAM run reported for the smaller Q3_K build. Build instructions target CUDA 13 on Linux and Windows.
License: what you can and can't do
Allowed: personal use, research, modification, noncommercial sharing, internal use and commercial use of outputs.
Needs a separate Edge0 license: selling or bundling the model, running a commercial model service, or offering a commercial coding or office assistant built on it.
Our take
A model this capable on a single gaming GPU is a real milestone for private, offline work with no token bills. But "open" here comes with commercial strings, the benchmarks are self-reported, and the hardware bar is still high-end. Great for personal and research use today; check the license before building a product on it.
Edge0-145B alternatives
Frequently asked questions
Is there an app?
Yes, the Edge0 App v0.2.0, which Edge0 says works offline with no account.
Where are the weights?
On Hugging Face, with deployment guides in the GitHub repository.
As featured in World of AI
DeepSeek puts its whole app into the API (Nov 13, 2026)
Facts checked against Edge0's announcement and Hugging Face model card on November 13, 2026.
Want four tools like this every weeknight? Subscribe to World of AI for free. Browse more in our AI Tools directory.