Picture a developer at her kitchen table, asking Edge0-145B, a 145-billion-parameter AI model, to debug a piece of code. There is no cloud bill ticking in the background. No data leaving her laptop. No queue, no API key. The model just runs, right there, on the graphics card she already owns.
That is the promise Edge0 made on November 12, 2026, when it released Edge0-145B, an open-weight AI model it describes as the first of its size built for personal computers rather than data centres. Alongside it came a smaller sibling, Edge0-120B, plus the software needed to run them both.
The claim is bold, and the fine print matters just as much as the headline. So does the licence. But if even half of what Edge0 says holds up, the line between cloud AI and home AI just moved a long way.
What Edge0-145B actually is
Edge0-145B is a latent sparse mixture-of-experts model, which is a mouthful that hides a simple idea. The model holds 145 billion parameters in total, but only about 10.12 billion of them switch on for any single token it generates. The rest sit idle, waiting for their turn.
That design is what makes the whole thing possible. A dense model this size would need everything loaded at once, which is why models this large normally live in data centres. Edge0-145B spreads its knowledge across 48 layers and 512 experts per layer, activating just 10 routed experts plus one shared expert for each token.
The model is derived from Qwen3.8-Flash-Next, Alibaba’s model, as the Hugging Face model card states plainly. This is not a model built from scratch. It is a specialised descendant, trained natively at 4-bit precision for local execution rather than compressed after a cloud training run.
It offers up to 256K tokens of context, which is enough to hold a large codebase or a long document in one conversation. The weights ship in two builds on Hugging Face: Q3_K at 40.81 GiB and NVFP4 at 47.79 GiB.
Edge0 did not just publish weights. It released edge0-infer, the inference framework that runs the model, and edge0-strata, the deployment tooling, on GitHub. A desktop app, Edge0 Chat, ships in version 0.2.0 for Windows, with an OpenAI-compatible local server and support for the Anthropic Messages API built in.
The one-GPU claim, and what it really costs
The headline number is 29 GB. That is the peak memory figure Edge0 reports in its announcement, paired with decoding speeds above 80 tokens per second. On an RTX 5090 D, using the Q3_K build, Edge0 reports 121.2 tokens per second for agent decoding and a score of 86.5 on Terminal-Bench 2.1, according to FreshSynapse’s breakdown of the launch.
Those numbers are Edge0’s own measurements, and they should be read as such. No independent lab has verified them yet. The benchmarks in the announcement are internal evaluations, and the promised technical report with reproduction instructions has not yet appeared.
The model card’s own residency numbers tell a fuller story. For the NVFP4 build, the card lists 23.25 GiB of video memory plus 38.27 GiB of system RAM, about 61.5 GiB in total. So “one GPU” is true, but it is one GPU plus a large helping of system RAM.
The trick that makes it work is predictive expert loading. Instead of holding the whole model in memory, the system predicts which experts the next token will need and loads them just in time. The expert library of about 40 GB lives in system memory, with roughly 48 GB of vocabulary data on the SSD.
The hardware requirements are specific. You need Linux or Windows, an NVIDIA GPU with CUDA 13, compute capability 7.5 or newer, at least 16 GiB of video memory and 20 GiB of system memory. Edge0 names laptops with 48 GB of memory as target devices.
One thing to note before you plan around a Mac: macOS is not a supported target for the 145B and 120B models, despite announcement wording that mentions a MacBook Pro. The GitHub platform table lists no macOS support. The smaller 35B tier, released in September, is the one that runs on phones.
The licence trap: open-weight is not open-source
Here is where the story gets complicated. Edge0-145B is open-weight, not open-source, and the distinction carries real consequences for anyone planning to build a product on it.
The weights sit under the Edge0 Community Model License 1.0, layered on top of the Qwen licence. Personal use, research, modification, non-commercial sharing and internal business use are all permitted. So is commercial use of the model’s outputs, at no charge.
But selling or bundling the model itself, operating a commercial model service, or offering an independent commercial coding or office assistant built on it all require a separate written licence from Edge0. The framework is Apache 2.0, which is genuinely open, but the weights are not.
That is a narrower deal than the one Meituan offered with LongCat-2.5, which arrived under the MIT licence earlier this month and allows commercial use, modification and redistribution. Alibaba’s Qwen4-Max, by contrast, never left preview at all.
Anyone planning a product on Edge0-145B should read the licence first, not the headline. The gap between “you can download it” and “you can sell with it” is exactly where this model sits.
Why local AI keeps getting bigger
Edge0-145B did not arrive in a vacuum. It landed in the same week that llama.cpp reached version 1.0, giving local AI apps a stable API for the first time, and that Thinking Machines published the weights of Inkling 2 Small. The open-weights ecosystem is having a very good week.
The direction of travel is clear. A year ago, running a capable model locally meant settling for something small. Then came phone-scale models, then models that needed a single powerful GPU, and now a 145-billion-parameter model that claims to run on a consumer card.
The appeal is easy to understand. A local model means no per-token API fees, no account, and no data leaving your machine. For developers, that is not a luxury. It is the whole point — and it lets a student with a gaming laptop try agent work that used to need a cloud budget.
What to check before you download
If you are tempted to try Edge0-145B, a few practical checks come first. Start with the GitHub README, not the launch blog, because the README is where the real platform support and requirements live.
Check your hardware honestly. The 29 GB headline figure is a measured peak in Edge0’s own setup, not a minimum. The model card’s residency numbers suggest you will want a well-equipped machine with plenty of system RAM, not just a strong GPU.
And treat the benchmarks as a starting point, not a verdict. Every performance number so far comes from Edge0 itself. Independent tests from the community will tell us far more in the coming weeks, and they are the numbers worth waiting for.
What happens next
The next test for Edge0-145B is not another announcement. It is the community. Independent developers will run the model on their own hardware, publish their own speeds, and find out whether the 29 GB figure survives contact with real machines.
Edge0 has promised a technical report with test setups and reproduction instructions. That document, when it appears, will be the real measure of how seriously to take the benchmark claims. Until then, the honest summary is that this is a promising, well-engineered release with unverified numbers.
Whether it becomes a tool developers actually rely on will depend on two things: whether the performance holds up outside Edge0’s own lab, and whether the licence lets the people who want to build on it actually do so.
For now, the model is there to download, the framework is there to run it, and a 145-billion-parameter brain can, in principle, live on a desk instead of in a data centre. That alone is worth paying attention to.



