AI Briefing#moe
The latest AI and developer news about #moe, with the original source and a short summary.
The latest AI and developer news about #moe, with the original source and a short summary.
FreeToken provides an edge-native serving system that efficiently runs large MoE models by leveraging the heterogeneous resources of personal devices.
The latest commit to the MS-Swift GitHub repository confirms support for the Qwen 3.8 35B MoE model.
Sand AI has released the MAGI-2 preview, a 114B parameter audio-visual MoE model.
Lumabri has been released, enabling distributed execution of MoE models in P2P swarm environments using the Colibri engine.
dots-studio has released a 280B MoE multimodal model supporting a 512K context window.
Nvidia reduced AI agent costs to 1/3 by routing to the optimal model for each task stage using Nemotron 3.5 Lightning and Switchyard.
inclusionAI has released an 8B MoE model with 1.3B active parameters.
Multimodal models developed more effectively when modalities were trained together from the beginning.
ANT Group's Ling 3.0 Tiny has been added to AI Gateway and is available for free until August 13.
Thinking Machines Lab released an open-weight model one-quarter the size of Inkling.
AMD has released a fully open 16B MoE model trained on ROCm.
AFM3 20B is a MoE model that selects only about 20% of MLP layers to activate per prompt.
MSLK has organized Attention, GEMM, and MoE kernel usage in one place.
Huawei has open-sourced openPangu-2.0-Pro, a 505B-parameter MoE model based on Ascend.
LLaDA2.2 has been released, overcoming the fixed-length output limitation of diffusion language models and improving agentic performance.
Molt is a PyTorch-based reinforcement learning framework that treats the agent itself as a program to support training of large-scale models.
Removing some low-weight routed experts in an MoE model was confirmed to have no impact on inference performance.
inclusionAI released LLaDA2.2-flash, an agent-oriented diffusion language model supporting 128K context and Levenshtein Editing.
Kwai released KAT-Coder-V2.5-Dev, an open-weight coding model with a 35B MoE architecture.
The Ascend NPU-based SLAI T-Rex framework achieved 34.22% MFU in post-training the trillion-parameter DeepSeek-V4 MoE model.
Poolside's Co-CEO Eiso Kant explains the training infrastructure and strategy behind building Laguna S, a 118B MoE model.
Upstage has released Solar Open 2, a 250B-scale open-weight MoE model that supports agentic workflows and 1M token context.
Soofi S, a highly efficient hybrid Mamba-MoE open-source model specialized for German and English, has been released.
xHC has been proposed as a technique that expands the Residual Stream into multiple streams to increase a model's memory capacity and intelligence.