Skip to content

Mixture Of Experts

Topic archive • 2 matches

AI-generated: summaries written by AI from the linked sources. How we use AI

Back to home • GEO summary endpoint

2026-11-11

Technology

  • Huawei detailed how its Kirin 9050 chip runs a 30-billion-parameter mixture-of-experts model on a phone without an internet connection, activating about 2 billion parameters at a time. The chip's Da Vinci NPU uses low-bit quantization and coordinates memory and compute to run the model locally.

    Hardware • TechNode

    Permalink

2026-11-14

Technology

  • Nvidia published a guide for serving GLM-5.3, a quantized 753-billion-parameter mixture-of-experts model from Z.ai with 40 billion parameters activated per token, across two DGX Stations connected via ConnectX-8 RDMA. The setup uses vLLM tensor parallelism to split weights and computation across both GPUs through a single endpoint, reaching 607.8 tokens per second.

    Infrastructure • Hacker News

    Permalink