A 145B-parameter open model built to fit on a 48GB laptop

Edge0 released two open models and says its 145B one runs on a laptop in about 29GB of memory, though all benchmarks are its own.

On November 12, 2026, Edge0 open-sourced two models, Edge0-145B and Edge0-120B, along with an inference framework and deployment tools. Edge0 calls the 145B model the first open-source on-device model at that size. Only about 10B of its parameters are active at each inference step.

On a laptop, Edge0 reports 29GB peak memory, more than 1,100 tokens per second for prefill, over 80 tokens per second for decode, about 120W typical power draw, and a 128K to 256K context window. The design groups experts in sets of 64 that share two large matrices, with each expert keeping a small set of its own parameters. Loading is tiered: hot experts sit in GPU memory, a roughly 40GB expert library in system memory, and about 48GB of vocabulary data on SSD, read on demand. Work is split between GPU and CPU, and speculative decoding checks several candidate tokens at once.

On Edge0's benchmarks, the 145B model scores 90.4 on GPQA-Diamond, 86.52 on Terminal-Bench 2.1 and 64.6 on DeepSWE 1.1. Its six-benchmark mean is 72.8, against 71.7 for Qwen3.8-Flash-Next and 70.1 for DeepSeek-V4-Flash-0731. Edge0 estimates those rivals need about 70GB and 87GB at 4-bit, so its model uses roughly 40% of the Qwen figure.

Why it matters: if near-frontier capability runs on hardware a person already owns, access to capable AI depends less on paid cloud APIs. Edge0 itself notes that hardware and electricity still cost money.

Caveats are significant. All scores come from Edge0's internal evaluations, and the rival memory figures are estimates, not measurements. The evaluation setup and reproduction steps are promised only in a later technical report. The text also gives 28GB in one place for the laptop memory figure. Independent testing is the thing to watch.

Source: Hacker News