Mistral Large 4: 1.05T parameters, ~400GB of RAM for local runs

Mistral has announced Large 4, an open-weight MoE model with 1.05 trillion total and 52 billion active parameters, but no weights are available yet — the Hugging Face page points to October 31. Running it locally takes roughly 400GB of fast memory at 2-bit quantization and about 617GB at 4-bit.

• 1.05T total parameters, 52B active, 1M-token API context
• API pricing: $1.36 per 1M input tokens, $4.18 per 1M output
• Estimated size: ~617GB at Q4, ~400GB at Q2
• License unnamed; weights promised by end of month