NVIDIA's new Vera Rubin platform, combining NVL72 and Groq 3 LPX, enables running agentic workloads on massive MoE models without sacrificing latency.
The hf-mem tool has added a detailed breakdown of memory consumption for Mixture-of-Experts (MoE) models, helping developers optimize their infrastructure strategies.
JetBrains has introduced Mellum2, a new generation AI model utilizing the Mixture-of-Experts (MoE) architecture with 12 billion parameters, specifically optimized for software development tasks and deeply integrated into IDEs.
Liquid AI introduces LFM2.5-8B-A1B, an 8-billion parameter language model with a hybrid MoE architecture, designed specifically for smartphones, laptops, and robots. Featuring a 128K context window, this is a major step forward in bringing high-performance AI directly to edge devices.
poolside has announced the Laguna M.1 and XS.2 model duo based on the Mixture-of-Experts architecture, optimized for coding tasks and long-horizon agents.
MiniMax has launched the M2 MoE model series with 229.9 billion parameters, optimized for agents and capable of self-debugging its own source code.
A new technique leveraging the CLC work-stealing mechanism enables CUDA Graph compatibility for grouped_gemm implementations, optimizing computational performance for complex AI models.
