Subscribe
0%
Illustration for: Edge0-145B 'runs on your laptop': what the repo actually says about GPUs, Macs, memory and the licence

Edge0-145B 'runs on your laptop': what the repo actually says about GPUs, Macs, memory and the licence

Edge0's new 145-billion-parameter open model is being shared as frontier AI on a laptop. Its own GitHub repo is narrower. The released build targets one NVIDIA GPU on Linux or Windows, Mac support isn't released, the model is derived from Qwen, and commercial hosting needs separate permission.

Listen to this article0:00 / 4:57

Edge0 released two open-weight models on November 12–13, Edge0-145B and Edge0-120B. Coverage of the launch has repeated the company's pitch: frontier-class AI on your own computer, a 29GB memory peak, more than 80 tokens a second, and day-one support on macOS, Windows and Linux, according to LavX News and Emergent.

The project's own GitHub repository describes something narrower, and readers trying to decide whether to download 40GB of weights should know the difference. Here's what the repo says, based on its README, model page and licence files as of November 14.

What was actually released

Both models are "derived from Qwen3.8-Flash-Next" and trained natively at 4 bits, the README says. They're sparse mixture-of-experts models: each token goes to 10 of 512 experts per layer plus one shared expert, so only part of the model does work at a time.

Edge0-145B Edge0-120B
Active parameters per token 10.12B 8.64B
Layers 48 48
Q3_K weights on disk 40.81 GiB 27.98 GiB
NVFP4 weights on disk 47.79 GiB 32.96 GiB
Engine edge0-strata, a fork of Strata same

The framework repo is Apache-2.0 and had about 10,950 stars on November 14, according to GitHub's search API. It first appeared in September with smaller 35B and 8B preview models.

Claim vs repo

"Runs on a laptop / macOS." The repo's platform table lists the 145B and 120B models for Linux and Windows with an NVIDIA GPU and CUDA 13 only. The macOS, iOS and Android columns are blank for these models. Those platforms apply to the older 35B model. The README includes an Apple M2 Ultra speed test but says this "does not establish macOS release availability." Emergent, for example, still says the model runs on personal computers generally.

"29GB peak memory." The repo's own measurements tell a fuller story. On an RTX 5090 D with 24GB of VRAM, the 145B Q3_K build was held as 23.14 GiB in VRAM plus 31.28 GiB in system RAM. In a constrained test capped at 16 GiB of VRAM and 20 GiB of RAM, with the rest streamed from SSD, it still ran, but more slowly. The repo calls that budget "a test limit, not a minimum system requirement."

"80+ tokens per second." Speeds depend heavily on hardware. The repo reports these figures for the 145B Q3_K build:

Machine Chat decode (tok/s) Agent decode (tok/s)
RTX 5090 D 24GB 99.7 121.2
RTX 4090 24GB 69.9 91.7
RTX Spark N1X 60GiB 37.7 36.7
Apple M2 Ultra (Metal, not released) 60.9 57.0
16GiB VRAM + 20GiB RAM cap 36.8 44.9

"Frontier capabilities." Every benchmark is Edge0's own. On its six-test table, the 145B model leads on DeepSWE 1.1 (64.6) and Terminal-Bench 2.1 (86.5). It trails its own Qwen base model on GPQA-Diamond (90.4 vs 91.7), IFBench (80.6 vs 81.3) and HLE (35.4 vs 35.9), and trails the listed Claude-Opus-4.6 score on HLE (40.0). DEV Community's explainer notes that nobody outside the company had run the model when it was written.

"Open source" comes with conditions

The code and the weights are licensed differently. The framework is Apache-2.0, and the edge0-strata engine keeps Strata's MIT licence. The model weights fall under a new Edge0 Community Model License 1.0 layered on top of Qwen's licence.

That licence allows personal use, research, benchmarking, internal business use, fine-tuning and redistribution, plus commercial use of the model's outputs. It reserves two things that need a separate written licence from Edge0:

  • selling the model, its weights or adapters, or bundling them as the main thing a customer pays for;
  • running a paid inference or fine-tuning service for third parties, such as an API or a hosted chatbot.

The licence also says Edge0's permission isn't Qwen's permission. Qwen's own commercial conditions still apply, and the repo's notice says no verified training-source commit is asserted for the release.

Who should try it

If you have a recent 24GB NVIDIA card on Linux or Windows, plenty of system RAM and a fast SSD, Edge0-145B is one of the larger models you can run on a single consumer GPU. The repo's edge0-infer serve command exposes OpenAI- and Anthropic-compatible endpoints. Mac owners should wait for a macOS release. Anyone planning to build a paid hosted product on it should read both licences first. Edge0 has promised a technical report with reproduction instructions, according to LavX, and independent benchmarks are the next thing to watch.

Sources

  1. GitHub — Edge0-AI/Edge0 README (145B / 120B release) github.com
  2. GitHub — Edge0-AI/Edge0: Edge0 145B and Edge0 120B model page github.com
  3. GitHub — Edge0-AI/Edge0: Edge0 Community Model License 1.0 github.com
  4. LavX News — Edge0 releases open-source 145-billion-parameter model for local inference on laptops news.lavx.hu
  5. Emergent — Edge0-145B: open-source 145B on-device model at 29GB emergent.sh
  6. DEV Community — A 145B open model that runs on your laptop: Edge0-145B explained dev.to

#Edge0#Edge0-145B#Edge0-120B#open-weight models#local AI#mixture of experts#Qwen#NVIDIA#licensing