Community · Nov 5, 2026, 20:47 UTC

Benchmarks show RustingStrata beating llama.cpp on large Qwen models on a 12 GB RTX 3060

In a community test on an RTX 3060 12 GB with 32 GB RAM, RustingStrata was about 1.5x faster on prefill and up to about 2x faster on decode for models larger than VRAM, while llama.cpp stayed ahead on small dense models. Numbers are from a desktop with other apps open and no tuning, and the author…

• LabAlibaba
• ModelQwen3.5-2B
• ModelQwen3-4B-Instruct-2507
• ModelQwen3.8-27B
• ModelQwen3-30B-A3B-Instruct-2507
• ModelQwen3.6-35B-A3B
• ModelQwen3-Next-80B-A3B-Instruct
• ModelQwen3.8-Flash-Next
• Topicbenchmark
• Topicopen-weights
• Topicchips