Skip to content- Sources: primary, discussion
- Summary: Edge0 released the weights of Edge0-145B and Edge0-120B with its inference framework. About 10B parameters are active per step. Hot experts stay in GPU memory, a roughly 40GB expert library sits in system RAM, and about 48GB of vocabulary data is read from SSD on demand. Speculative decoding spreads the cost of weight reads. Edge0 reports 29GB peak memory, more than 1,100 tokens per second prefill and 64.6 on DeepSWE 1.1, all from internal evaluations.
- Why it matters: If independent runs confirm the numbers, a 48GB laptop could run a coding agent model competitive with hosted flash-tier APIs, with no token fees.
- Comments: The Hacker News thread drew 221 points and 60 comments. The benchmark comparisons are Edge0's own, and the full evaluation setup is promised in a technical report.