Can I run Devstral 2 123B?

Devstral 2 123B is Mistral's big dense coding model, and it is squarely for people building serious local coding setups, not the laptop crowd. At 4-bit it weighs about 74.6 GB, which already tells you the story: it does not fit on a 12 GB RTX 3060 or a 24 GB RTX 4090, full stop. You need roughly 96 GB of memory to load it comfortably, so realistically that means a high-memory Apple Silicon machine or a workstation with serious RAM. This is a model you plan your hardware around, not one you casually pull down to try.

In daily use the honest caveat is speed. On an Apple M Max you are looking at around 5 tokens per second, and a CPU-only DDR5 box drops to roughly 1 token per second, which is closer to batch-job territory than interactive chat. It reads and writes code well, but you will feel every reply stream in slowly. The context window is a generous 256K on paper, though memory is the real limit: pushing to 128K context already wants about 132 GB total, so on a 96 GB machine you keep working context modest and lean on shorter, focused prompts.

Against the MoE Qwen 3.5 122B-A10B in related_models, the trade-off is clear: that model activates only a slice of its weights per token and generally feels faster at a similar parameter count, while Devstral 2 runs every one of its 123B parameters on each token. Devstral's standout is being a focused, dense coding specialist from Mistral with a long context, if you have the memory to feed it. One practical note on the license: it ships under a Modified MIT license, so read the specific terms before any commercial deployment rather than assuming plain MIT freedom.