Edge0-145B Brings a 145-Billion-Parameter Coding Model to Your Own Machine

Most big coding models live in someone else’s data centre. You send your code off, you wait, you pay by the token, and you hope the pricing page doesn’t change next month. Edge0 is pitching a different deal. The company has announced Edge0-145B and its smaller sibling Edge0-120B, a pair of downloadable models built for coding, reasoning and agent-style tasks that run on your own machine instead of in the cloud.
The headline number is the parameter count. Edge0-145B carries 145 billion parameters in total, which sounds like something that should need a server rack and a small power station. Yet the reported peak memory use is around 29 GB, and the reported speed is over 80 tokens per second on compatible NVIDIA hardware. How does a model that size move that fast on local kit? The answer sits in how the model is built.
A Sparse Model, Explained Plainly
Edge0-145B is a sparse mixture-of-experts model. Think of a normal dense model as one very large brain that lights up completely for every single word it processes. A mixture-of-experts model works more like a newsroom full of specialists. There’s a large pool of expert sub-networks, and a router sends each piece of work to only the few experts best suited to it.
So while the full model holds 145 billion parameters, only about 10 billion of them are active for any given step. The rest sit idle for that token. That’s the whole trick, and it’s a good one. You get the knowledge stored across a huge model, but you only pay the computing cost of a much smaller one each time it generates.
Why does that matter for running locally? Because the active parameter count is what drives speed. A model that only fires up roughly 10 billion parameters per step can plausibly hit the reported 80-plus tokens per second on a strong consumer or workstation GPU setup, instead of crawling along at a pace that makes real coding work painful.
The Numbers That Matter at Your Desk
Let’s line the reported figures up. Total parameters: 145 billion. Active per step: around 10 billion. Peak memory: roughly 29 GB. Speed: 80-plus tokens per second on compatible NVIDIA hardware. Those numbers, if they hold up in real use, put the model in an interesting spot. 29 GB of memory is beyond a modest laptop, but it’s well within reach of a serious workstation or a machine built with AI work in mind.
There’s a privacy angle here too, and for coding work it’s not a small one. When a model runs on your own hardware, your source code never leaves the building. No prompts shipped to a third-party API, no proprietary codebase sitting in someone else’s logs. For freelancers handling client work and small teams guarding unreleased products, that’s a genuine reason to pay attention.
Speed matters just as much. An assistant that answers at a trickle breaks your flow; one that keeps up at 80 tokens a second feels closer to a colleague typing back at you. Agent tasks benefit even more, since an agent may chain dozens of steps together, and slow generation compounds at every step.
Two Sizes, One Idea
Edge0-145B isn’t launching alone. Edge0-120B arrives alongside it, following the same downloadable, sparse design for local coding, reasoning and agent tasks. Two sizes suggest Edge0 wants to cover more than one class of machine, giving people whose hardware can’t stretch to the bigger model a smaller door into the same approach.
I like seeing the pair announced together. A single giant model makes a splash; a family hints at a product line the company actually expects people to run day to day.
Read the Licence Before You Download
Now the honest caveat, and it’s a real one. The licensing on these models is restrictive, and it deserves scrutiny before anyone builds plans around them. “Downloadable” is not the same as “free to use however you like.” Restrictive terms can limit commercial use, redistribution, or how the model gets deployed in a product, and those details live in the licence text, not in the announcement.
So do the boring thing. Read the licence end to end before you integrate Edge0-145B into client work or a paid tool. If the terms don’t fit your use case, the smart move is to wait, ask, or walk away, no matter how appealing the specs look. Plenty of developers have been burned by skipping that step with other models.
Where to Look Next
If you want to explore the local-model scene further, Hugging Face remains the main hub where downloadable models get published and discussed, and NVIDIA’s site is the place to check whether your hardware falls into the “compatible” bracket Edge0 describes. For context on where the wider market is heading, our coverage of Claude Haiku 5.5 shows the cloud side pushing just as hard on speed, while the LangChain funding story shows how much money is flowing into agent tooling right now.
Edge0-145B’s pitch is simple: big-model knowledge, small-model running costs, your own hardware. The reported numbers back that pitch up on paper. Whether it earns a permanent place on developers’ machines will come down to real-world testing, and to whether that licence lets people actually use it the way they want.
