Dev.to AI 🤖 Ai 👁 0 📖 2 min read

A 145B Coding Model You Can Actually Run at Home

Cloud coding assistants come with a quiet bargain: your code travels to someone else's servers, and you pay for every token on the way back. Edge0 wants to cancel that bargain. The company has announced Edge0-145B, along

Cloud coding assistants come with a quiet bargain: your code travels to someone else's servers, and you pay for every token on the way back. Edge0 wants to cancel that bargain. The company has announced Edge0-145B, alongside a smaller Edge0-120B — downloadable models aimed at coding, reasoning and agent tasks, designed to run on your own machine.

Here's the puzzle. The model holds 145 billion parameters, yet Edge0 reports peak memory of about 29 GB and speeds above 80 tokens per second on compatible NVIDIA hardware. How?

The trick is sparsity. Edge0-145B uses a mixture-of-experts design, which you can picture as a newsroom of specialists rather than one giant brain. A router hands each step to just a few expert sub-networks, so only around 10 billion parameters are active at any moment. The rest stay idle. You keep the knowledge of a huge model while paying the compute bill of a far smaller one — and that active count is exactly what decides generation speed.

For developers, the local angle cuts two ways. First, privacy: your codebase never leaves your hardware, which matters when the code belongs to a client or an unreleased product. Second, flow: an assistant answering at 80-plus tokens a second keeps pace with your thinking, and agent workflows that chain dozens of steps don't grind down under slow generation.

29 GB of memory still rules out a modest laptop, mind you. This is workstation territory. But it's a very different proposition from renting a data centre by the token.

One warning before anyone gets excited: the licensing is restrictive. Downloadable doesn't mean do-as-you-please, and the terms deserve a careful read before this goes anywhere near commercial work. Specs are the fun part; the licence is the part that decides what you're actually allowed to build.

I covered the announcement in full, including the two-model lineup and the licence caveat, in my write-up: Edge0-145B Brings a 145-Billion-Parameter Coding Model to Your Own Machine.

If the reported numbers survive real-world testing, local models this capable stop being a hobbyist curiosity. They become a genuine option for working developers — licence permitting.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.