Edge0 has open-sourced a 145-billion-parameter AI model that it says runs on an ordinary laptop, a release that pushes the frontier of on-device AI far beyond what local hardware was thought to handle. The company announced Edge0-145B and a smaller sibling, Edge0-120B, on November 12, along with the inference framework and deployment tools needed to run them.
Until now, the trade-off in local AI has been brutal: capable models needed server-class hardware, while models small enough for a personal computer struggled with demanding work. Edge0 claims its release breaks that trade-off. The company calls Edge0-145B the world's first open-source on-device AI model at the 145-billion-parameter scale, and says it brings coding, reasoning and agent capabilities that rival frontier cloud models to a personal computer.
How a 145B model fits on a laptop
The trick is sparsity. Edge0-145B is a latent mixture-of-experts model built on Qwen3.8-Flash-Next, with 145 billion total parameters but only about 10.12 billion active for any given token. Because a mixture-of-experts model only uses a small slice of its parameters at each step, the rest can sit on disk and be loaded on demand, so peak memory is bounded by the active experts rather than the total parameter count.
On a compatible laptop, Edge0 reports peak memory of 29GB, decoding above 80 tokens per second, prefill above 1,100 tokens per second, roughly 120 watts of typical power draw, and a context window of 128,000 to 256,000 tokens. The company says the model can run on a 48GB laptop, such as a high-memory MacBook Pro. The weights are published on Hugging Face, the code is on GitHub under an Apache 2.0 licence, and the repository has already passed 10,000 stars.
The benchmark numbers — and the caveats
Edge0's own testing puts the model near or ahead of named cloud models on several benchmarks. It reports 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7; 90.4 on GPQA-Diamond, close to DeepSeek-V4-Flash at 90.8; and 86.52 on Terminal-Bench 2.1. Across six benchmarks it reports a mean of 72.8, against 71.7 for its Qwen baseline.
Those figures come from Edge0's internal evaluations, and no independent verification has been published yet. The comparisons are vendor-chosen, and some baseline figures are estimates rather than measurements, so they are a starting point for testing rather than a verdict. There are also practical limits: the 145B and 120B models need a Linux or Windows machine with an Nvidia GPU running CUDA 13, and 29GB of peak memory restricts the release to high-end machines. Edge0's phone tier is a separate 35B model.
Why local AI matters
The appeal of running a model locally is straightforward: no per-token API fees, no data leaving the machine, and tools that keep working offline. Edge0 frames the release against a history of constrained local models, noting that memory limits kept most on-device models below 10 billion parameters before September 2026.
If the performance claims hold up under independent testing, the release marks a genuine shift — frontier-scale reasoning moving from rented cloud GPUs onto hardware people already own. The code, weights and licence are real and public, so anyone with a compatible machine can now test that claim for themselves.



