Edge0-145B: A 145-Billion-Parameter AI Model That Runs on Your Laptop — Here's the Catch

On November 12, a small team called Edge0 open-sourced a 145-billion-parameter AI model that it says runs on a single consumer GPU at more than 80 tokens a second. I read the announcement, the model card, and the skeptical follow-ups — here's what the numbers really mean, what the license actually allows, and whether a WordPress developer should download it.

Edge0-145B: A 145-Billion-Parameter AI Model That Runs on Your Laptop — Here's the Catch
Edge0-145B: A 145-Billion-Parameter AI Model That Runs on Your Laptop — Here's the Catch — Technical Architecture & Overview

Table of Contents

The Headline That Stopped My Scroll

I have followed local AI for years, and the rule has always been simple: the big models live in data centres, and the models on your own machine are small. Anything above ten billion parameters was a cloud job — rented by the token, with your code travelling to somebody else's server.

So when a small team called Edge0 announced on November 12, 2026 that it had open-sourced a 145-billion-parameter model that runs on a single consumer GPU at more than 80 tokens a second, I stopped scrolling. A model that size doing real coding work on hardware you can own would change the economics of the AI tools I use every day — especially for people like me, who work on client WordPress sites and would rather not send client code to a cloud API. I read the announcement, the model card, and the skeptical follow-ups. The short version: this is a real release, the engineering is clever, and the headline number needs a footnote the size of a paragraph.

What Edge0 Actually Shipped on November 12

Edge0 released two models together — Edge0-145B and a smaller sibling, Edge0-120B — with its unified inference framework and deployment tools. The company calls Edge0-145B the world's first open-source on-device model at 145 billion parameters, with coding, reasoning, and agent capabilities rivalling frontier cloud models.

The numbers: 145 billion total parameters, only about 10.12 billion active per token. Forty-eight layers, 512 experts per layer, 10 routed experts plus one shared expert per token. A 256,000-token context window and native 4-bit training — built for low-precision local hardware, not squeezed down to it afterwards. Two builds are on Hugging Face, Edge0-145B-q3k and Edge0-145B-NVFP4, and the Edge0 App (v0.2.0) runs the models without a build environment.

One detail matters: the architecture is a latent mixture-of-experts design built on Qwen3.8-Flash-Next, Alibaba's roughly 180-billion-parameter model from August, according to the Hugging Face model card metadata. Edge0 emphasises training "in its on-device architecture from the outset," and both can be true — but if you pictured a model trained entirely from scratch by a small team, the model card tells a more ordinary story.

How a 145B Model Fits on One Machine

The trick is four ideas working together. First, the mixture-of-experts design: in a dense model every parameter works on every token, but here only about 10 billion of the 145 billion participate in any single step — each token is routed to a small set of specialist experts and the rest sit idle. That is why the total count sounds impossible for a laptop while the compute per token is merely large.

Second, the "latent" part. The routed experts work through a narrow width — 1,792 dimensions instead of the full 2,560 — and each group of 64 experts shares two large matrices, so every expert becomes a small fixed-size record of about 1.6 MB that can be pinned, swapped, or streamed onto the GPU as needed.

Third, the memory hierarchy. Hot experts stay in GPU memory, an expert library of roughly 40 GB sits in system memory, and about 48 GB of vocabulary data lives on the SSD, memory-mapped from disk and read one row per token. Nothing loads all at once; the engine streams what each token needs.

The 29GB Number — Read the Fine Print

Edge0's headline figure is 29 GB of peak memory with decoding above 80 tokens per second on a compatible personal laptop — a number that travelled everywhere within hours. Strictly speaking, it is only the GPU-side peak. The model card's own measurement for the NVFP4 build is 23.25 GiB of VRAM plus 38.27 GiB of host RAM — about 61.5 GiB in total — measured on an RTX 5090 D with 24 GB of VRAM — a vendor measurement, including cache and buffers. But the direction is clear: "29 GB" is the GPU peak, and the machine still needs tens of gigabytes of system RAM and the better part of 100 GB of fast disk.

There is a second wrinkle. Edge0 names 48 GB laptops — a MacBook Pro, for example — and 64 GB workstations as the target hardware. But the engine for the 145B and 120B models, edge0-strata, needs an NVIDIA GPU with CUDA 13 on Linux or Windows; macOS is not supported for these two models. The smaller edge0-35b tier runs on a GGML-based engine with Metal support, including iOS — that is the tier that actually runs on a Mac or a phone.

So does a 145-billion-parameter model run on your laptop? If yours is a 48 GB machine with a high-end NVIDIA GPU and a fast SSD, the honest answer is apparently yes, with caveats — and nobody independent has confirmed it yet. If yours is the 16 GB ultrabook most of us carry, the answer is no; the 35B tier is the model for you. A 145B-class model on one consumer GPU would have sounded absurd two years ago — but "runs on any laptop" is a different claim, and the distinction matters before you clear 100 GB of disk for it.

The Benchmarks — Impressive, and Entirely Theirs

On DeepSWE 1.1, a benchmark of real software-engineering work, Edge0-145B scores 64.6, ahead of its Qwen base model at 58.7 and DeepSeek-V4-Flash at 54.4. On a terminal-use test it reports 86.52, and across a six-benchmark mean, 72.8 against 71.7 for the Qwen base. Edge0 also says the model plans, writes, and runs code, debugs errors, and checks results visually — for 3D scenes, games, hardware programs, and PCB designs.

The catch: every score comes from Edge0's own internal evaluations. The technical report with the setup and reproduction instructions is not out yet, and no independent runs have been found. Until somebody outside the company reproduces the numbers, the table is a claim, not a settled measurement. That does not make the numbers false — the DeepSWE lead is large enough to be interesting, and the weights and code are public, so verification can start now.

The Licence — Open, With Three Locked Doors

The framework code is Apache 2.0 — genuinely open. The weights use the Edge0 Community Model License 1.0 with the Qwen Community License, which reserves three uses for a separate written licence: selling, renting, or licensing the weights; operating a model service for third parties financed by fees, subscriptions, advertising, or sponsorship where customers control inference inputs; and offering an independent AI coding or office-productivity product financed the same way, hosted or on-device.

In plain language: download it, run it, use it for your own work — fine. Build a paid coding assistant on top of it and sell that assistant, or run a hosted service where customers bring their own prompts — that needs a conversation with Edge0 first. A freelancer using it to write code faster is unaffected; a founder planning a product on it should read the licence before writing the business plan. "Open weights" and "do anything you like commercially" are not the same thing.

What This Means for WordPress and Elementor Developers

Three reasons, all practical. First, privacy. Every prompt to a cloud AI coding tool contains somebody else's code, business logic, sometimes customer data. A capable model on your own machine removes that problem entirely — your code never leaves the building.

Second, cost. Cloud coding agents charge by the token, and agentic workflows are token-hungry — they read whole codebases, run, fail, retry, and read again. Local inference has no per-token bill. Hardware and electricity still cost money, but the marginal cost of one more agent run drops to near zero.

Third, the agent wave is already inside WordPress. I covered it recently: Elementor's MCP support, Breakdance 3.0 and Oxygen 6.2 shipping agent connectors, Divi's AI agent. Those agents need a brain, and until now it has almost always been a cloud model. A local 145B model that can plan, write, run, and debug code is a different kind of brain — one that lives on your machine, costs nothing per run, and never lets client data leave the office. Pair it with the edge0-infer OpenAI-compatible endpoint and the shape of a fully local WordPress workflow starts to appear.

The honest caveat: roughly 10 billion active parameters is not a frontier cloud model, so the realistic setup is hybrid — local for everyday work, cloud for the hardest problems. But the local side of that hybrid just got dramatically more capable, and that is the real story here.

Should You Download It Today?

If you have a 24 GB NVIDIA GPU, 64 GB of RAM, and a fast SSD with room to spare — and you enjoy experiments — yes. Treat the first run as a weekend experiment, not a production dependency. If you are on a Mac or a 16 GB laptop, wait: the 145B tier is not built for your machine yet, and the 35B tier is the on-device option for you today. If you need reliable production coding help this week, keep your cloud subscription — the benchmarks are vendor-run, the technical report is pending, and a model you cannot yet verify is not one to bet a deadline on. And read the licence before you build a product on it.

Conclusion

Edge0-145B is a real release with real engineering behind it: a 145-billion-parameter latent mixture-of-experts model, about 10 billion parameters active per token, that its makers say runs on one consumer GPU at more than 80 tokens a second. The design — shared matrices, a narrow latent width, hot experts on the GPU, a memory-mapped expert library and vocabulary on disk — is exactly the kind of systems thinking that moves local AI forward. The coding benchmarks, if they hold up, put it ahead of its base model and in the conversation with serious coding tools.

But the headline needs its footnotes. Twenty-nine gigabytes is the GPU-side peak; the model card measures about 61.5 GiB in total. The MacBook in the marketing is not yet matched by the CUDA-only engine. Every benchmark is vendor-run, the technical report is still promised, and the licence keeps three commercial doors locked. None of that makes the release uninteresting — it makes it one to try with open eyes. Two years ago, a model this size was a data-centre job. Now it is a download, a command line, and a machine you can own. For WordPress developers, for freelancers with client code to protect, and for anyone tired of paying per token, that direction is the news. I will be watching for the independent benchmarks — if they confirm even half of what Edge0 claims, the way we build websites is about to get a lot more local.

Frequently Asked Questions

What is Edge0-145B?

Edge0-145B is an open-weight AI model released on November 12, 2026: 145 billion total parameters, about 10.12 billion active per token, built on Qwen3.8-Flash-Next as a latent mixture-of-experts model. It runs coding, reasoning, and agent work on a single consumer NVIDIA GPU instead of in the cloud. The weights are on Hugging Face as Edge0/Edge0-145B-q3k or Edge0/Edge0-145B-NVFP4.

How much memory does Edge0-145B really need?

Edge0 reports 29 GB of peak memory, but that is the GPU-side peak. The model card's own measurement for the NVFP4 build is 23.25 GiB of VRAM plus 38.27 GiB of host RAM — about 61.5 GiB in total — plus disk space for the roughly 40 GB expert library and 48 GB of vocabulary data.

Can I run Edge0-145B on a Mac?

Not the 145B or 120B tier, for now. The edge0-strata engine needs an NVIDIA GPU with CUDA 13 on Linux or Windows, and macOS is not a supported target for these models. The smaller edge0-35b tier runs on a GGML-based engine with Metal support, including iOS.

Is Edge0-145B really open source?

Partly. The inference framework is Apache 2.0, but the weights use the Edge0 Community Model License 1.0 plus the Qwen Community License. Personal use is fine, but selling the weights, running a paid hosted service, or selling an AI coding product built on it all need a separate written licence.

How good is it at coding?

On Edge0's own benchmarks it scores 64.6 on DeepSWE 1.1, ahead of its Qwen base at 58.7. All vendor-run so far — no independent verification yet.

Looking for more WordPress & Elementor tutorials?

Explore our complete directory of 53+ web performance, theme development, and Elementor architecture guides.

View Topic Sitemap XML Feed