Moe is a topic tracked in our intelligence system with 6 linked articles.
TechCrunch publishes a living AI glossary with concise, practical definitions of key terms (e.g., AGI, LLM, RLHF) and notes its ongoing updates, plus a small event promo embedded in the page.
A 2016 Intel Xeon server with 128 GB DDR3 RAM and no GPU runs a 26B Mixture-of-Experts model using CPU-optimized inference and a long, flag-heavy tuning process, illustrating memory-bandwidth limits and the claimed viability of open-weight AI on commodity hardware.
A research paper demonstrates Rotary GPU enabling local execution of large Mixture-of-Experts models on consumer hardware (8 GB VRAM), achieving 2048 tokens at ~6.3 GB VRAM and ~21 tokens/sec, signaling edge-deployment viability under VRAM constraints.
Liquid AI unveils LFM2.5-8B-A1B, an 8B parameter MoE edge model with 128K context, 38T pretraining, expanded tokenizer, and strong on-device benchmarking and tool-calling capabilities.
Kog claims real-time LLM inference on standard datacenter GPUs can reach about 3,000 tokens/s per request on a 2B model by co-designing a monokernel runtime, GPU code, and a Laneformer architecture, with scalability toward frontier MoEs as memory bandwidth grows.
ZAYA1-8B is a sub-1B active-parameter open-source MoE model (8.4B total) trained entirely on AMD hardware, achieving competitive math/coding benchmarks and highlighting an AMD-focused pathway with open weights and proprietary inference tech.
Subscribe for real-time topic updates and unlimited access to our intelligence platform.