Qwen 72B runs on RTX 5090 at 115 tokens/sec, BF16 weights 55.6GB

AccBalanced · x · 2026-08-18

Demonstrates running Qwen 2.5 72B (possibly mislabeled as 3.8-27B in original tweet, likely 72B or 32B given the context) locally on an RTX 5090 (32GB VRAM) system, achieving 115 tokens/sec. Notes that the official BF16 checkpoint for Qwen 2.5 72B is 55.6 GB.

Original post →

More from Models

Models channel →