A little-known company says its large AI model runs on one gaming PC with no cloud fees. Only its own tests back that up

A little-known AI company called Edge0 says it has fitted a 145-billion-parameter AI model onto a single consumer graphics card. If that holds up, a company could run a capable coding assistant on one desktop PC. It would no longer pay a cloud provider for every word.

Edge0 announced the model, Edge0-145B, on November 12 alongside a smaller 120B sibling. Release notes followed on November 13. By Saturday its most popular compressed version showed more than 14,000 downloads on Hugging Face, the main site for sharing AI models.

The pitch targets cloud AI pricing directly. Edge0's launch post says running locally means no per-use fees to a cloud AI provider, and that data can stay on the computer. It also concedes that hardware and electricity still cost money.

The catch: nearly every number comes from Edge0 itself. No independent tester had published results by Saturday, and several of the company's claims conflict with its own documentation.

What Edge0 claims

Parameters are the internal settings a model learns. More usually means more capable, and more memory. Edge0's model holds 145 billion but uses only about 10 billion for each word fragment it produces. The rest wait in system memory or on the storage drive and load when needed.

That design is what lets a large model squeeze onto one graphics card. Edge0 did not build the model from scratch, though. Its model card says it is derived from Alibaba's Qwen3.8-Flash-Next, reworked to need far less memory.

Edge0 estimates the Alibaba original needs about 70GB in a standard compressed form. It says its own version peaked at 29GB in its measurements.

On six tests it chose, Edge0 reports an average score of 72.8. That is slightly above its Alibaba parent at 71.7 and DeepSeek's V4 Flash at 70.1. Its clearest gains come on two tests of AI agents doing coding work: 64.6 versus 58.7 on DeepSWE 1.1, and 86.5 versus 83.3 on Terminal-Bench 2.1.

It trails its parent on three of the other four tests, including a graduate-level science quiz and a very hard general-knowledge exam, and edges it on the fourth by 0.2 points. In other words, the reworking mostly preserved the original's ability rather than surpassing it.

Edge0 also shows itself beating Anthropic's Claude Opus 4.6 on Terminal-Bench, 86.5 to 78.2. That comparison flatters it. Anthropic has since replaced Opus 4.6 with newer models, and even the older Claude still leads on the hardest exam, 40.0 to 35.4.

On speed, Edge0's model card reports about 100 tokens a second in chat on an Nvidia RTX 5090 D card with 24GB of memory. Tokens are the word fragments models read and write. Edge0 says Nvidia's older RTX 4090 managed about 70.

Where the claims don't line up

The launch post says the model runs on a laptop such as a MacBook Pro. But Edge0's own GitHub documentation lists Macs as unsupported for the 145B and 120B models, according to reviews by eesel and Zeniteq. Those models require an Nvidia card on Linux or Windows.

The model card itself says its results on an Apple machine do not establish that a Mac version is available.

The memory figures also diverge. Edge0's model card says its full-speed setups used about 23GB of graphics memory plus 31GB to 38GB of system memory. That is roughly double the 29GB headline.

The downloaded files themselves run 41 to 48GB. In an Edge0 test capped at 16GB of graphics memory and 20GB of system memory, chat speed fell to about 37 tokens a second.

The smaller sibling has inconsistencies too. Edge0's 120B scores differ between its launch post and its GitHub page, eesel found. Edge0 says full test methods will come in a technical report that had not appeared by Saturday.

Commentary on the launch carries its own interests. Eesel sells AI customer-support software. MindStudio, which sells a platform for building AI apps, advised companies to test the model on their own work before relying on it.

What buyers should weigh

The license matters as much as the speed. Edge0 calls the model open-source, but its community license is narrower. It allows personal, research and internal company use, plus commercial use of what the model produces.

Some uses need a separate written license from Edge0. These include selling or bundling the model, or running a paid AI service on it. So does offering a commercial coding or office assistant. Alibaba's Qwen license terms apply as well. Edge0 has published no price for the commercial license.

For a company that wants an in-house coding helper on its own hardware, with data never leaving the building, those terms may be fine. For a software vendor hoping to undercut cloud AI prices, the model is a negotiation with Edge0, not a free option.

Edge0 also has a thin public profile. Its launch post says it has released about 200 models and counts more than 5 million devices and 110,000 customer organizations. None of the outside coverage reviewed for this article confirms those figures or names the company's backers.

The broader signal is the direction of travel. Large models are being reworked to run on hardware a business already owns, which puts pressure on per-token cloud pricing for routine work. Until outside testers reproduce Edge0's numbers, treat them as a vendor's opening claim.