Edge0-145B: a 145B AI model that runs on your laptop
Edge0-145B is a new open-weight model that runs offline in about 29GB of memory. What hardware you need, how fast it is and how it compares.
Checked and updated 13 November 2026 at 03:30 UTC

Key points
- Edge0-145B is a 145-billion-parameter open-weight model released on November 12, 2026 that runs on a laptop in about 29GB of peak memory.
- Only about 10B parameters are active per step, which is how it reaches 80+ tokens per second on supported Macs and Nvidia PCs.
- It is free to download from Hugging Face, but you need a high-memory machine; benchmark wins are Edge0's own tests.
Edge0-145B is a new open-weight AI model, released on November 12, 2026, that its makers say brings near-frontier performance to a personal laptop: it needs about 29GB of memory at peak and runs offline at more than 80 tokens per second. You can download it free from Hugging Face, but you need a powerful, high-memory machine to run it.
What it is
The model has 145 billion parameters, but it is a "mixture of experts": only about 10 billion are active for each step, which keeps it fast and small enough for a laptop. Hugging Face lists Qwen3.8-Flash-Next as its base model and the weights in a 4-bit (NVFP4) format. Edge0 released a smaller Edge0-120B, its own inference framework and a desktop app at the same time.
What you need to run it
| Edge0's figures | |
|---|---|
| Memory | About 29GB peak (one line says 28GB) |
| Macs | Apple M-series, e.g. MacBook Pro |
| PCs | Nvidia RTX 4090, RTX 5090, RTX Spark |
| Systems | macOS, Windows, Linux from day one |
| Power | About 120W typical |
| Speed | 80+ tokens/s writing, 1,100+ tokens/s reading |
| Context | 128K to 256K tokens |
If your computer has 16GB or less, this model is not for you. Edge0 says low-power devices need a smaller model or the cloud; it quotes a separate Edge0-35B running on a phone at about 2.5GB.
How good is it?
Edge0's own tests put it close to much bigger cloud models on coding and reasoning. A few of its numbers:
| Benchmark | Edge0-145B | Comparison (Edge0's table) |
|---|---|---|
| Terminal-Bench 2.1 | 86.5 | Claude 78.2 |
| DeepSWE 1.1 | 64.6 | Qwen3.8-Flash-Next 58.7 |
| GPQA-Diamond | 90.4 | Claude 91.3 |
| HLE | 35.4 | Claude Opus 4.6 (Max) 40.0 |
These are internal evaluations, some baselines come from other sources, and the full setup is promised in a later technical report. Treat them as a claim until independent testers check them.
How to try it
- Check your machine has well over 29GB of free memory.
- Download the Edge0 app (v0.2.0) from Edge0's GitHub, or the weights from Hugging Face (Edge0/Edge0-145B-NVFP4).
- Use the one-click deployment and hardware guides that come with it.
The app's core features need no account and no usage quota.
Should a normal user care?
Running AI locally means your prompts never leave your computer and there's no monthly bill. The trade-off is hardware cost, and no live web knowledge without the cloud. Our guide to on-device vs cloud AI covers that choice, and AI models that run on your phone lists lighter options.
AI Tools This Week is not affiliated with Edge0.
Frequently asked questions
Is Edge0-145B free?
There is no price. The weights are on Hugging Face and the Edge0 app is on GitHub. Running it locally has no token or API fees, but you pay for the hardware and power. Hugging Face lists the licence as "other", so read it before commercial use.
What computer do I need to run Edge0-145B?
Edge0 says it needs about 29GB of memory at peak and runs on Apple M-series Macs and Nvidia RTX 4090, RTX 5090 and RTX Spark machines, on macOS, Windows and Linux. It mentions a MacBook Pro and a 48GB laptop as examples.
Does Edge0-145B work offline?
The core features of the Edge0 app work offline with no account or usage quota. Edge0 says things like cross-device memory and searches that need current information still need the cloud.
Sources
Related
Haru-4B: the open AI voice model that talks and listens at once
Haru-4B from Japan's higurashi is a free, open voice model that listens and speaks at the same time, in 9 languages. What it does and how to run it.
Idol image-to-3D model: what it is and is it free?
Speridlabs' Idol turns one photo into a 3D Gaussian splat, including the sides you can't see. Free under Apache 2.0, but it needs an NVIDIA GPU.
Qwen3.8-27B: what it is, is it free and can you run it?
Alibaba's Qwen3.8-27B is a free, open AI model that runs on a strong home PC. What it does, file sizes, how to install it and benchmarks.