How to Run Edge0-145B Locally: The 145B Coding Model That Fits in 29GB

Edge0-145B is a 145-billion-parameter open-weight language model from a startup called Edge0 that's built specifically to run on a single consumer machine instead of a GPU cluster. It shipped on November 12, 2026, alongside a smaller Edge0-120B sibling, and the pitch is simple: a sparse mixture-of-experts design keeps only about 10 billion parameters active per token, so peak memory lands around 29GB instead of the 300GB+ you'd normally need for a model this size. It also speaks an Anthropic-compatible API out of the box, so you can point Claude Code at it on localhost. I spent a few hours going through the launch materials, the Hacker News thread, and the actual GitHub repo, and the short version is: it's real, it's clever, and the marketing page is doing some rounding that you should know about before you buy hardware for it.

What Edge0-145B actually is

Edge0-145B is built on top of Qwen3.8-Flash-Next, not trained from scratch. Edge0 took that base, applied a sparse MoE setup, and tuned the whole thing for 4-bit local execution (NVFP4 quantization, if you care about the format). The inference engine, called edge0-strata, is a fork of an existing open-source project named Strata (a few people on Hacker News pointed this out within hours of launch), and Edge0 doesn't really hide it either.

Here's the spec sheet as Edge0states it:

A smaller Edge0-120B shipped the same day for people who want a bit more headroom, and Edge0 also maintains a separate, much smaller on-device model line for phones. Don't confuse the two: the 145B is the laptop/desktop flagship.

The memory math doesn't quite add up

The headline number everywhere is 29GB, and that's the figure Edge0 leads with on its own blog. It's technically true, but it's also not the whole bill. Dig into the fine print and you find Edge0 also wants roughly 40GB of system RAM for the expert-routing layer plus 48GB of SSD storage for weight streaming. Add it up and you're closer to 70GB of real footprint, not 29GB. A few commenters on the Hacker News thread did exactly this math and called it out. Fair, because "runs in 29GB" and "needs 70GB of resources to actually run" are different claims.

To be clear, 29GB peak VRAM/unified memory for inference itself checks out against the benchmarks Edge0 published. It's the surrounding system requirements that got left out of the headline.

There's a second gap worth flagging: the launch page talks up macOS support, and the MacBook Pro-with-48GB-RAM numbers are literally what's in the spec table above. But the actual code currently released on GitHub only supports CUDA 13 on Linux and Windows. One Hacker News commenter put it bluntly: "there's no guidance on how to run on a mac at all." So if you were planning to try this on Apple Silicon this week, you can't, despite what the marketing implies. That may change with a future release, but as of today it hasn't.

How to actually run it

If you've got an NVIDIA GPU with CUDA 13 and at least 16GB of VRAM (some third-party writeups say you'll want 24GB to be comfortable, and I'd plan for the higher number), here's the path:

1. Pull the weights. They're on Hugging Face under Edge0/Edge0-145B-q3k.
2. Get the inference engine. Clone Edge0-AI/edge0 from GitHub, or skip the custom build entirely and use llama.cpp: llama serve -hf Edge0/Edge0-145B-q3k.
3. Or use the native binary. edge0-infer serve --model /path/to/Edge0-145B-q3k spins up a local server with OpenAI- and Anthropic-compatible endpoints.
4. Point your agent at it. Set your tool's base URL (for Claude Code, that's ANTHROPIC_BASE_URL) to the local server address instead of Anthropic's API, and your existing coding-agent workflow should talk to Edge0-145B without any code changes on the client side.

That last part is the genuinely useful bit. A lot of "local model" releases make you rewrite your tool-calling logic to match some custom schema. Edge0 instead cloned the Messages API shape closely enough that Anthropic-format clients work with just a URL swap: no account, no API key, no network call once the model's downloaded.

Does it actually beat Claude Opus 4.6?

Edge0's own numbers say yes, on one benchmark: Terminal-Bench 2.1, where they report 86.52 for Edge0-145B against 78.2 for Claude Opus 4.6. The rest of the published scorecard is DeepSWE 1.1 at 64.6, GPQA-Diamond at 90.4, IFBench at 80.6, and a mean of 72.8 across six benchmarks.

I'd hold off on the victory lap. These are Edge0's own internal evaluations: nobody outside the company has reproduced them yet, and benchmark gaming through targeted fine-tuning is a real pattern in this industry, not a hypothetical. The Hacker News thread had its share of skeptics on exactly this point. On the practical side, real-world reports are mixed: one user running a circuit-board routing task said the model "would have to significantly alter the placements" to be usable, and another doing Python interview-style problems said it "never produced a coherent answer" even after 8,000+ tokens of back-and-forth. That's not damning on its own (every model has rough edges on niche tasks), but it's a reminder that a strong Terminal-Bench score doesn't automatically translate into reliable day-to-day coding help.

Open-weight, not open-source

Edge0's own copy calls this "the world's first open-source on-device AI model with 145 billion parameters." That phrase is doing a lot of work it shouldn't. The weights ship under the Edge0 Community Model License 1.0, plus the Qwen Community License that comes along for the ride since it's built on Qwen3.8-Flash-Next. Neither is OSI-approved open source. If you want to sell access to the model, bundle it into a commercial product, or stand up a competing coding assistant service with it, you need a separate written license from Edge0. The inference code on GitHub is genuinely Apache 2.0, which is nice, but the weights (the actual expensive, hard-to-reproduce part) aren't.

This matters if you're a solo developer running it on your own machine for your own code: you're almost certainly fine under the community license. It matters a lot more if you work at a company thinking about shipping a product on top of it.

Should you bother?

If you've got a CUDA 13 box with 24GB+ of VRAM sitting around, yes, it's worth trying: it's free, it runs fully offline, and the Anthropic-compatible local API means you can test it against your actual Claude Code workflows in about twenty minutes. Just go in knowing the real resource ask is closer to 70GB total than the 29GB headline, Mac support isn't here yet despite the marketing, and the benchmark numbers are vendor-reported until someone independent reruns them. For teams with code that can't leave the building, that's still a real reason to kick the tires. For everyone else chasing the "beats Claude Opus 4.6" headline, I'd wait for third-party numbers before planning around it.

For more on running capable coding models outside the big labs' clouds, see our breakdowns of Ripple 65B running on a MacBook and self-hosting Kimi K3.1 Pro's API. And if you want the baseline Edge0-145B is actually built on, check out Qwen 3.8 Max's release.

Watch and read more