Edge0 releases an open 145B model that operates on a laptop

The Edge0 team released Edge0-145B and Edge0-120B as open-source models, with the tools to operate them. Edge0-145B has 145 billion parameters (the values in a model). Only about 10 billion work on each step. The team reports that it operates on a laptop and is near some cloud models on coding tests. All results come from tests by the team.

How it works

The model has many experts (small parts of the model). Groups of experts use the same large matrices (grids of numbers). This makes the data to keep and move less. The team made the model for local use from the start.

Frequently used experts stay in GPU memory. The other experts stay in system memory and on the SSD. The CPU and the GPU work together.

The numbers

• Terminal-Bench 2.1: 86.52, against 78.2 for Claude.
• DeepSWE 1.1: 64.6, more than Qwen3.8-Flash-Next.
• GPQA-Diamond: 90.4, near the 91.3 of Claude-Opus-4.6.

The team did these tests. No external test is given.

What is not known

The report gives no name for the laptop. The model must have 28 GB of memory, which is more than a phone can give. The memory and speed numbers come from different models on different hardware.

Insight: The key point moves from model size to the software

Amazon has an NVIDIA model with 120 billion parameters that uses only 12 billion for each token. Google made one graph that operates on CPU, GPU and NPU hardware. Models that use a small number of parameters at a time use less memory. This reads as a move of the key point to the software that moves weights on the device. Builders must test that software, not only the model.