Betelgeuse runs 4-bit MoE at 24 tok/s on RTX 3060 with 16GB RAM, 100% byte-exact

zyxciss · reddit · 2026-11-01

The Betelgeuse author addressed two criticisms from a previous thread — that IQ2XS was too low-bit, and that testing on a single 16GB DDR4 stick was unrepresentative — by benchmarking the full 4-bit IQ4XS (100GB-class experts) on the same RTX 3060 12GB under both RAM configs with 16GB total.

Key results:

Why dual-channel doubled throughput: Betelgeuse uses an exclusive host-RAM victim cache plus async hardware DMA (MOEDMA=1) between the VRAM slot pool and NVMe. Single-channel DDR4 (17-19 GB/s) forces CPU and PCIe DMA to fight over one memory channel; doubling bandwidth lets the GPU's DMA engine pull victim-cache hits at full speed, jumping 14 → 24 tok/s.

Engine on GitHub, pre-packed bundles on HuggingFace, and an upstream llama.cpp PR are live. Next: porting the same streaming + victim-cache architecture to DeepSeek V4.1 Flash, already hitting 25 tok/s on similar hardware in early testing.

Related event: Betelgeuse runs 68GB Qwen MoE on RTX 3060, benchmark data released(2 posts)→

Original post →

More from Infra

Infra channel →