On November 12, 2026, Edge0 announced on its official blog that it is open-sourcing Edge0-145B and the smaller Edge0-120B, together with its inference framework and deployment tools. Edge0-145B is billed as the world's first open-source 145-billion-parameter on-device model, meant to move coding, reasoning, and agent work from the cloud onto personal computers. The code is released under the Apache 2.0 license, and the weights are on Hugging Face.
How do you fit 145B parameters into a laptop?
The trick is that only a small part runs at a time: 145B is the total parameter count, but only about 10 billion take part in each inference step. Sixty-four experts share two large matrices, each keeping only a small set of its own parameters. Loading is split across three tiers too: frequently used experts stay in GPU memory, an expert library of about 40GB sits in system memory, and about 48GB of vocabulary data stays on the SSD to be read on demand; the CPU and GPU work in parallel, and speculative decoding verifies several candidate tokens at once. The headline figures are 29GB peak memory, over 1,100 tokens per second of prefill, over 80 tokens per second of decoding, 128K–256K context, and about 120W typical power, with day-one support for macOS, Windows, and Linux, including Apple M-series chips and NVIDIA RTX platforms. These numbers come from Edge0's own internal testing and have not been independently reproduced.
Benchmarks close to cloud models — with the caveats in plain sight
On Edge0's internal evaluations, Edge0-145B scores 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7; 86.52 on Terminal-Bench 2.1 and 90.4 on GPQA-Diamond, with a 72.8 mean across six benchmarks and memory use of roughly 40% of a comparable baseline. In its demos the model did serious work locally on its own: building a Manhattan street scene in Blender, writing the program that let a real drone find a red backpack, and drawing a circuit board from scratch in KiCad before turning it into an assembly animation — all without a cloud call. The bar for local models has not disappeared, though: 29GB peak memory means an ordinary 16GB laptop cannot run it; 48GB machines are the target. That matches the conclusion when Mistral Large 4 opened its 1T-parameter weights: downloadable weights do not mean everyone can run them. Earlier local tools such as Ollama made small models easy to use; Edge0 is betting on the other end — making the big model itself smaller and cheaper to run, then fitting it into a personal machine.
For developers, the real change is not the benchmark scores but how the bill works: local inference has no per-token charge, code and data can stay on the machine, and work continues offline; the price is hardware and electricity you pay yourself, while model updates, long-term memory, and checking the latest information still need the cloud. Whether the 145B can repeat the buzz of Edge0-35B, which topped Hugging Face's trending chart when it launched in September, depends on whether community testing holds up the official numbers.
