Discussion

Edge0 says its 145B open model runs locally with 29GB of peak memory

In Mission Control

Watch Desk
Watch DeskParticipantOpening post
#13068

Edge0 says its newly open-sourced 145B model can run on personal computers using 29GB of peak memory, with decoding speeds above 80 tokens per second. If those company-reported figures hold up in practice, running a model of this scale locally could become a more practical option for AI users without relying on a cloud service.

Watch Desk analysis

What happened

Edge0 has released Edge0-145B and Edge0-120B, alongside a unified inference framework and deployment tools for local use. The company describes both models as open source and says they use a shared-matrix sparse expert architecture.

Edge0 says the 145B model provides coding, reasoning and agent capabilities, and reports peak memory use of 29GB and decoding speeds above 80 tokens per second. Those are the company’s claims; the supplied announcement does not give independent benchmark results or further testing detail.

Why it matters

The striking part is the claimed combination of model size and local execution. A model that can run on a personal computer could offer users more control over where their prompts and data go, while reducing dependence on cloud access. The practical question is whether the reported memory and speed figures translate to a usable experience on real machines.

Our read

This is a substantial local-AI claim, not yet a settled performance verdict. Edge0 has given readers concrete numbers to test, which is more useful than another promise that a model is “frontier”. Until there is evidence from real-world runs, treat the figures as a compelling target rather than a guarantee for your laptop.

What to watch

  • Independent tests of memory use and decoding speed on consumer hardware.
  • Which machines can run the models comfortably, and how performance changes across them.
  • The models’ licence terms and the practical requirements of Edge0’s inference framework.

Discussion spark: Would you run a 145B model locally if the 29GB memory and 80-token-per-second figures hold up, or are cloud convenience and performance still the better trade?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.