Mistral Large 4: A 128B Open-Weight Model Under Apache 2.0

Looking for a frontier-class model you can run on your own hardware without asking a vendor for permission? Mistral Large 4 is a 128B dense model with a full Apache 2.0 license on Hugging Face.

That means no commercial-use restrictions and no guardrails baked in by the vendor. The trade-off is the hardware bill: this is not a laptop model.

What Happened

Mistral Large 4 is a 128B dense model, published under Apache 2.0 on Hugging Face. It has a native 256k token context window, no stiff vendor guardrails, and no commercial license restrictions.

The hardware the source lists for running it:

• VRAM: 2x 48GB cards (A6000 Ada or A40), or 4x RTX 3090/4090 24GB cards for Q4/FP4 quantization.
• System RAM: 128GB to 256GB.
• Storage: around 260GB for the FP16 weights, or around 70GB for the Q4_K_M quantized version.

Why It Matters

If you want a frontier model with a transparent license instead of an opaque one, Apache 2.0 removes the usual legal question. You can put this in a commercial product without negotiating terms, which is rare at the 128B size.

The catch is cost. You need two 48GB cards, or four 24GB consumer cards if you quantize, plus a machine carrying 128GB to 256GB of RAM. Quantizing to Q4_K_M brings the download down to roughly 70GB, but you still need the same class of hardware underneath it. Plan the machine before you plan the deployment.