#moe
Moe: 22 AI articles covering moe news, analysis, and research
Articles
Introducing Olmo-core 3: Open, Scalable Training Infrastructure⭐9
AI2's Olmo-core 3 delivers open, scalable training infrastructure for large Mixture-of-Experts models, tackling expert parallelism, load balancing, and checkpoi...
StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE⭐8
StepFun's Step 5 Preview is a 600B-parameter MoE model with 27B active parameters and 1M context, built for long-horizon agentic work like coding and finance.
Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With⭐8
Jina AI releases jina-ocr-v1, a 3.4B MoE document parser with built-in speculative decoding, scoring 91.14 on OmniDocBench and built for low-budget GPUs.
DeepSeek-V4.1-Flash Launches with 1M Context, FP4 KV Cache, and⭐9
DeepSeek-V4.1-Flash launches with 1M context, FP4 KV cache, and cross-layer attention reuse—cutting KV memory to 890 bytes per token for input-heavy agent workl...
GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Chinese AI Labs⭐8
Two Chinese labs independently release similar frontier MoE models: GLM-5.3-Flash and Qwen3.8-Flash-Next, sharing a hybrid attention recipe.
From Mixtral to Kimi K3: The Evolution of Mixture-of-Experts Models⭐8
From Mixtral to Kimi K3, explore Mixture-of-Experts evolution from 8 to 896 experts per layer, compression, and stability for trillions of parameters.
Meet FreeToken: An Edge-Native MoE Serving Engine that Runs 753B⭐7
FreeToken enables running massive 753B-parameter MoE models on a single workstation GPU, democratizing high-scale AI inference for individual developers.
NVIDIA Unveils Nemotron 3.5 Lightning: A 30B Open MoE with 3B⭐8
NVIDIA launches Nemotron 3.5 Lightning, a 30B open MoE with 3B active params, plus NeMo Switchyard router for faster, efficient AI agents.
Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic⭐7
Cursor open-sources MoK, a deterministic MoE training megakernel for NVIDIA GB300 NVL72 racks. It boosts throughput up to 2.37x for Composer models.
Alibaba Qwen Unveils Qwen3.8-Max: A 2.4-Trillion-Parameter MoE⭐6
Alibaba unveils Qwen3.8-Max, a 2.4T-parameter MoE model with open weights next week, plus Qwen3.8-27B for practical deployment.