AI Briefing#mixture-of-experts
The latest AI and developer news about #mixture-of-experts, with the original source and a short summary.
The latest AI and developer news about #mixture-of-experts, with the original source and a short summary.
NVIDIA highlights performance per watt as the key metric determining the profitability of AI infrastructure in power-constrained environments, showcasing the overwhelming efficiency of the Blackwell platform.
Gemma 4 is a next-generation open-weight multimodal model with diverse architectures and innovative features.
We propose PathMoE, which constrains the expert paths that tokens traverse to improve the efficiency and performance of MoE models.
Poolside has released Laguna XS 2.1, an MoE model optimized for agentic coding.
LongCat-2.0, a large-scale MoE-structured model with a total of 1.6 trillion parameters, has been released.
As LLM architectures grow more sophisticated, a shift in design approach is needed to secure both performance optimization and research flexibility at the same time.
Laguna-M.1, a 225B-parameter MoE model optimized for agentic coding, has been released.
The NVIDIA Blackwell platform recorded overwhelming performance across every category of the MLPerf Training 6.0 benchmark, proving its leadership in AI training infrastructure.
NVIDIA has optimized Google DeepMind's new diffusion-based model DiffusionGemma to run up to 4x faster on RTX and DGX systems.
Google has unveiled DiffusionGemma, an experimental model that moves away from the sequential generation approach of existing LLMs and generates blocks of text simultaneously.
Cohere has released North Mini Code, an MoE model optimized for agentic software engineering.
Cohere has released a new coding model with 30B parameters as early access via Hugging Face.
Nvidia's MoE reasoning model **Nemotron 3 Ultra** has launched on Vercel AI Gateway, supporting agentic workflow optimization.
JetBrains has unveiled **Mellum 2**, a 12B-parameter MoE model specialized for software engineering.
JetBrains has unveiled Mellum2, a 12B Mixture-of-Experts (MoE) model.
Using OpenAI Triton, a high-performance MoE inference kernel supporting both NVIDIA and AMD has been developed.
Cohere has open-sourced Command A+, a 218B-parameter MoE model, under the Apache 2.0 license.
Ovis2.6-80B-A3B has been released as an MoE-based multimodal model.
Ai2 released EMO, showing that performance can be largely preserved using only a subset of a MoE's experts.
Mistral AI is rapidly expanding into the enterprise market with a sovereign open-weight strategy.
A PR has been submitted to llama.cpp adding support for the Sarvam-specific MoE architecture.
EMO has been released, maintaining performance even with only some experts used, thanks to document-level routing.
SpecMD benchmarks MoE expert cache policies and proposes Least-Stale.