Command R+ Unaligned 104B Needs Four 24GB GPUs

Running a 104B model at home is not a laptop job, and Command R+ Unaligned is no exception. This one is built for a very specific kind of work, and it asks for a small rack of hardware to do it.

It is a large open-weight model tuned for pulling accurate answers out of your own documents, so if that is what you are shopping for, here is what it costs to run.

What Happened

Command R+ (Unaligned 104B) is a 104B parameter open-weight model optimized specifically for enterprise-scale RAG, or Retrieval-Augmented Generation. The unaligned variant is the part worth noticing: it keeps its document citation accuracy while not cutting sensitive corporate data out of the answers.

The hardware it expects for Q4/Q5 quantization:

• VRAM: 4x 24GB, or 2x 48GB cards such as the A6000.
• System RAM: 128GB.
• Storage: around 60GB for the GGUF quant.
• Context window: 128k tokens.

The context window is the quiet headline here. 128k tokens is enough to hand it a long report, or several, without chopping them into pieces first, which is exactly the point of a RAG-tuned model.

Why It Matters

Most people hearing "open-weight" picture a model that runs on one consumer card. This one does not. Two 48GB A6000s, or four 24GB cards, plus 128GB of system RAM, puts it in workstation or small server territory, and the 60GB of disk means you should plan the space before you start the download.

If you already have that class of machine, the trade is straightforward: a 104B model with a 128k context that is tuned to quote your documents back to you accurately, without screening out the material you fed it.