Edge0-145B: a 145B Model That Runs on Your Laptop

Edge0 just did the thing every local-AI fan has been waiting for someone to actually do. On 12 November 2026 the company launched Edge0-145B, a 145-billion-parameter open-weight model it claims is the world's first open-source 145B model built to run on a device. Not a datacentre. Not a rented GPU cluster. Your laptop.

I have not run it yet — nobody outside Edge0 really has, it is day one — so this is a launch-day read, not a hands-on review. But the numbers, if even half of them survive contact with real hardware, change what "local AI" means.

What actually launched

Two models, actually: Edge0-145B and its smaller sibling Edge0-120B. The headline one is 145B total parameters in a sparse mixture-of-experts design, with only around 10B parameters active per token. That is the trick that makes the whole thing plausible — a 145B model that only lights up a tenth of itself for each token is a very different beast from a dense 145B model, and it is why the memory and speed numbers below are not fantasy.

It is derived from Qwen3.8-Flash-Next and trained specifically for 4-bit local execution. The weights are on Hugging Face (Edge0/Edge0-145B-q3k), the code is on GitHub under Apache 2.0 (Edge0-AI/edge0), and the repo has already passed 10,000 stars. macOS, Windows and Linux from day one. No account, no cloud call, fully offline.

For context: Edge0's previous release was Edge0-35B in September 2026, aimed at phones. This is the same playbook scaled up — take a strong open model, quantise and tune it hard for local hardware, ship the weights. The launch post hit 410K views on X and the Hacker News front page, which tells you how hungry people are for this exact thing.

The numbers that matter

All of these come from Edge0 itself, and I will keep saying that, because it matters:

• Runs on a MacBook Pro with 48GB of memory. Peak memory 29GB, decode at 80+ tokens per second, prefill at 1,100+ tokens per second, drawing around 120W. That is a usable, fast local model on a machine a lot of developers already own.
• Context of 128K to 256K tokens. Long enough for real work — big codebases, long documents, multi-file reasoning — without the window closing halfway through the job.
• Benchmarks, vendor's own: an average of 72.8 across six benchmarks. On DeepSWE 1.1 it scores 64.6, against 58.7 for its parent Qwen3.8-Flash-Next and 54.4 for DeepSeek-V4-Flash. On Terminal-Bench 2.1 it scores 86.52. On a hard science test it scores 90.4, slightly below a rival's 91.3 — Edge0 admits that one, which is a good sign.

The DeepSWE number is the one I would watch. Beating the model it was derived from by nearly 6 points on agentic coding is a real claim, and it is also the easiest kind of claim to check independently once people start running the weights. The Terminal-Bench score is the same story: if it holds, this is a genuinely capable coding model that lives on your desk.

Why this is a big deal, if it holds

The honest version of the local-AI story so far has been a compromise: you could run a small model privately, or a big model in the cloud. The small ones were fine for chat and summaries and fell over on hard coding. The big ones needed an API key, a bill, and your data leaving your machine.

A 145B model that runs at 80 tokens a second on a laptop, offline, with no account, collapses that compromise. It means agentic coding work — the kind of multi-step, tool-using work that used to need a frontier API — running on hardware you already own, on code you never have to upload. For anyone working on private code, client code, or anything under NDA, that is not a nice-to-have. It is the whole point.

It also matters for cost. A model that runs locally has no per-token bill. Run it all day, run it on a hundred files, run it while you sleep — the marginal cost is electricity. That changes which jobs are worth automating.

The catch, and it is a real one

Here is the part the launch headline does not say. According to Edge0's own repo docs, Linux and Windows need an NVIDIA GPU with CUDA 13 and at least 16GB of GPU memory. So "runs on your laptop" has a footnote, and the footnote is doing a lot of work: on Linux and Windows, your laptop needs a serious discrete GPU. The MacBook numbers are the clean ones — Apple Silicon's unified memory is what makes the 29GB peak fit — and that is exactly the machine most developers do not have in a Windows-heavy office.

So read the launch claim carefully. It is a 145B model that runs on a laptop, not on any laptop. If you are on a MacBook Pro with 48GB, the numbers above are aimed at you. If you are on a thin Windows ultrabook with integrated graphics, this is not your model yet, and the repo docs say so plainly. Credit to Edge0 for putting that in the docs rather than burying it.

The other catch is the usual one: every number in this article is Edge0's own. No independent lab has run these benchmarks yet, no one has verified the 80 tokens per second outside Edge0's test setup, and launch-day charts are picked to look good. The 90.4 on hard science, slightly below a rival's 91.3, is at least an admission that the model does not win everything — vendors that admit a loss somewhere are usually telling you something real. But the rest needs independent runs, and they will come fast now that the weights are public.

Who should try it, and who should wait

Try it this weekend if: you have a MacBook Pro with 48GB (or a Linux/Windows machine with a 16GB+ NVIDIA GPU and CUDA 13), and you do coding work you would rather keep off the cloud. The weights are free, the code is Apache 2.0, there is no account to create. Download it, run the same coding task you ran on an API model last week, and compare the output — not the benchmark chart, your task.

Try it this week if: you are building anything where data cannot leave the machine — client work, internal tools, offline environments. A 145B model with a 128K+ context window that never phones home is a genuinely new building block for that kind of product.

Wait if: your machine does not meet the GPU bar, or your work is the hard-science corner where Edge0 itself concedes a rival is ahead. Also wait before rebuilding any production pipeline around the benchmark scores — DeepSWE and Terminal-Bench numbers from the vendor are a reason to test, not a reason to migrate.

Do not switch on the benchmark chart alone. Every launch looks like a win on its own chart. The honest test is your hardware, your prompts, your codebase.

The honest caveats

• All benchmarks and speed figures are Edge0's own. DeepSWE 1.1 (64.6), Terminal-Bench 2.1 (86.52), the 72.8 average, the 80+ tokens per second, the 1,100+ prefill — all vendor claims on day one, with no independent verification yet.
• The GPU requirement limits the headline. Linux and Windows need an NVIDIA GPU with CUDA 13 and at least 16GB of GPU memory, per the repo docs. The "runs on your laptop" claim is real for a MacBook Pro with 48GB and for well-specced machines; it is not universal.
• Derived, not built from scratch. Edge0-145B is derived from Qwen3.8-Flash-Next, so its strengths and its blind spots will rhyme with that family. The improvement over the parent on DeepSWE is the interesting part — and the part most worth verifying independently.
• No hands-on testing was performed for this article. Everything here comes from Edge0's announcement (12 November 2026) as published on edge0.ai, the Hugging Face repo and the GitHub repo.

Edge0-145B is not the biggest model of 2026, and it is not trying to be. It is trying to be the biggest model that fits on your desk. If the numbers hold up outside Edge0's lab, that is a bigger shift than another point on a benchmark chart — it is the point where frontier-class coding stops being something you rent and starts being something you own.

Have you tried running it yet? If you have a machine that meets the bar, I would genuinely like to hear whether the decode speed and the coding quality match what Edge0 is claiming — that is the part no launch post can prove.

Sources: Edge0 announcement, "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer" (edge0.ai, 12 November 2026); Edge0/Edge0-145B-q3k on Hugging Face; Edge0-AI/edge0 on GitHub (Apache 2.0, 10,000+ stars). All benchmark, speed, memory and power figures are Edge0's own claims; no independent verification exists yet and no hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.