Edge0-145B: A 145B AI Model That Runs on a 48GB Laptop
In short
- Edge0 open-sourced Edge0-145B, a 145B-parameter model that runs on a laptop with 48GB of memory.
- Reported peak memory is 29GB and speed is 80+ tokens per second, with no cloud token fees.
- If you pay for API tokens on agent work, this is the first local model worth testing against that bill.
Until now, a big model meant a big cloud bill. Edge0 says that is changing. On November 12 the company released Edge0-145B and Edge0-120B, both with weights on Hugging Face.
What happened
Edge0-145B has 145 billion parameters, but only about 10 billion work on each step. That is how it fits. Edge0 reports 29GB of peak memory, 80+ tokens per second when generating, 1,100+ tokens per second when reading a prompt, and a 128K to 256K context window. Power draw is around 120W.
The benchmarks are strong on paper. On Terminal-Bench 2.1 it scores 86.52 against 78.2 for Claude Opus 4.6 (Max). On GPQA-Diamond it gets 90.4, close to Claude’s 91.3. On DeepSWE 1.1, a coding test, it scores 64.6.
One warning. These are Edge0’s own internal evaluations, and the full setup comes later with a technical report. Treat the numbers as a claim, not a verdict.
It runs on macOS, Windows and Linux, including Apple M-series chips and NVIDIA RTX 4090 and 5090 cards.
How to use it
The best fit is agent work: coding, debugging, running scripts, checking results in loops. Those jobs eat thousands of tokens per task. Local means every extra round costs you nothing but electricity.
It is also the answer for private files. Contracts, customer lists, board layouts can stay on your machine.
Do the math: say you spend €150 a month on API tokens for agent work. That is €1,800 a year. Running Edge0-145B for 8 hours a day on 22 workdays at 120W is about 21 kWh. At €0.25 per kWh that is roughly €5 a month. The catch is the laptop. You need 48GB of memory, and Edge0 gives no price for anything else, so the saving depends on hardware you may already own.
What it means
Cloud labs lose first on repetitive, high-volume work. Anyone who runs the same agent task all day now has a reason to move it local and keep the cloud for the hardest jobs.
Before you build on it, read the license. I did not verify the terms for commercial use, and that matters if you plan to ship a product.
Read also: Run a 27B Qwen Model on Your Laptop for €0 With LM Studio · NVIDIA’s Free Bolt Models Run AI Agents on Your Own PC · Anthropic Now Spends More on AI Than on Salaries
Ivan’s Take
Worth a test, not a switch. If I had a 48GB machine, I would move my repetitive agent jobs there this week and compare the quality against my cloud bill. I would wait for the technical report before I trust any benchmark.
Source: Edge0
HawkValue
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
