Open-weight AI models you can run on your own hardware

Open-weight models keep arriving, and they matter for a simple reason: you can run them on your own hardware instead of sending client data to a vendor's cloud.

On October 29, Mistral published the weights for Mistral Large 4, a roughly 1-trillion-parameter multimodal model with about 52 billion active parameters and a 1-million-token context.

Two weeks earlier, Meta released Muse Spark Open (816 billion total, 42 billion active).

NVIDIA also shipped Nemotron 3.5 Super VL (120 billion total, 12 billion active) and Nemotron 3.5 Bolt, built to run everyday agent work locally in 24 GB. Liquid AI's LFM3, out October 27, is a 1.7-billion-parameter text model that takes about 1 GB at 4-bit, so it can run on a laptop or a phone.

That range is the story for a small business. Keep sensitive work on-prem, control cost, and stay online if a hosted API goes down or refuses a legitimate security task.

Read the licence first. Mistral's is a modified MIT: fine for most commercial use, but a Model-as-a-Service offering needs a separate licence. Meta's Muse Model License has limited commercial terms. The big models need serious hardware. And Mistral itself warned of "a potential for uplift to conduct real attacks." If you host one, lock down access, log what it does, and keep it off the open internet.

Originally published at https://trevorheshka.com/blog/open-weight-ai-models-you-can-run-on-your-own-hardware/