Skip to main content
Search

# MoE serving

2 related articles

Subscribe to updates
  1. Gaming PCs as one elastic inference platform: FreeToken targets 290B+ MoE models on consumer hardware

    FlashML released FreeToken, an edge-native Mixture-of-Experts serving engine aimed at running frontier open-weight MoE models above 290B parameters locally on personal and consumer hardware. It treats GPUs, CPUs, host memory and interconnects as one elastic inference platform, organizing compute and data movement through bandwidth-adaptive CPU–GPU co-execution, full-layer double-buffered prefill streaming and a global LRU expert cache. The engine also adds semantic anchor checkpoints to skip recomputation after agentic context edits and reallocates VRAM at runtime between expert caches and KV memory, covering models such as DeepSeek-V4-Flash, Qwen3.6-35B-A3B and GLM-5.2 across MXFP4, NVFP4, FP8 and BF16 formats.

  2. arXiv

    MoE precision pays twice: store one 4.27 bpw copy, read 2.9 bpw per workload, and stay near 4-bit formats

    Mixture-of-experts weight precision carries two separate costs, one for storage and one for decode; DuoRate stores a single 4.27 bpw copy and picks a read prefix per workload from that workload's own prefill, reaching 76.16 average accuracy at a 2.9 bpw read rate on Qwen3-30B-A3B, close to 4-bit formats.

You’re all caught up