Betelgeuse runs 68GB Qwen MoE on RTX 3060, benchmark data released

A llama.cpp fork called Betelgeuse runs a 68GB Qwen MoE model at full precision on an RTX 3060 with 16GB RAM at 22 tok/s. The author later addressed criticism over low-bit quantization and atypical memory tests, releasing upgraded victim-cache benchmarks showing 24 tok/s with 4-bit MoE.

2026-10-31 ~ 2026-11-01 · 2 related posts

• Betelgeuse runs 68GB Qwen MoE bit-exact at 22 tok/s on RTX 3060 + 16GB RAM — zyxciss · 2026-10-31
• Betelgeuse runs 4-bit MoE at 24 tok/s on RTX 3060 with 16GB RAM, 100% byte-exact — zyxciss · 2026-11-01