The Trillion Parameter Model Fits on Four GPUs, in a Bite

Mistral put the weights for a 1 trillion parameter model on Hugging Face on 29 October, and the serving maths that came with it matters more than the benchmark scores.

• Mistral Large 4 carries 1 trillion parameters with 49 billion active per token and a 1M token context window.
• vLLM shipped day-0 support. The NVFP4 checkpoint serves at tensor parallel 4 on Blackwell, the FP8 checkpoint at tensor parallel 8.
• Decode context parallelism lifts KV cache capacity about 7.9 times across eight GPUs.
• The drafter built into the model accepts about 2.9 tokens per step, roughly doubling output throughput on eight GPUs.
• Mistral claims 47 percent fewer tokens per task than GLM-5.3, and lists $0.70 per million input tokens against $2.10 per million output.

The capacity is bought with GPUs, not created by the algorithm. The cache, not the parameter count, sets how many users one replica holds. Read the full breakdown.