This week in Training Methods · updated Thu Nov 12 2026
This week in model research, advancements in training and inference efficiency took center stage. Meta FAIR introduced e-means, a faster k-means variant that achieves comparable clustering performance in just 1.4 to 2.6 epochs—dramatically reducing training time—and is now integrated into FAISS 1.16.0 with up to 45x GPU speedups. Simultaneously, deeper insights emerged into Mixture of Experts (MoE) routing mechanisms, exploring how models like Mixtral and DeepSeek-V4 assign tokens via Top-K softmax, sigmoid gating, and hash-based strategies, each balancing performance and computational cost. On the hardware optimization front, Zyphra released open-source FP8 attention kernels for AMD MI355X, delivering up to 1.76x faster inference than native implementations while reducing error rates by 62–78%. These developments highlight progress in algorithmic efficiency, routing intelligence, and hardware-specific acceleration.
AI-written weekly briefing drawn from this topic's recent stories.
Meta FAIR has introduced e-means, an algorithm that outperforms traditional k-means by reaching comparable loss levels in just 1.4 to 2.6 epochs instead of 25. Integrated into FAISS 1.16.0, it delivers 10-45x faster GPU …
This article examines various routing mechanisms in Mixture of Experts (MoE) architectures, detailing how different models assign input tokens to specialized expert networks. It covers strategies like Top-K softmax, sigm…
Zyphra has released open-source FP8 and BF16 attention kernels tailored for AMD Instinct MI355X, achieving up to 1.76x faster performance than AMD’s native AITER implementation. The new kernels maintain high accuracy, wi…
Together AI has introduced Together Compiler, a system that uses autonomous agents to continuously optimize inference performance in real time as workloads change. The approach achieved up to 2.6 times higher throughput …
NVIDIA demonstrated three architectural and software-level optimizations that increase inference throughput for mixture-of-experts (MoE) models on its DGX Spark platform. When applied to Ant Group's 124B Ling-3.0-flash m…
Anthropic has launched Message Threads in its Claude platform, reducing time to first token by up to 35% for lengthy interactions by transmitting only changes since the last message. The feature improves efficiency for a…
qwenGPU is a specialized version of Qwen3.5-2B that integrates 45 compiled GPU layers directly into its architecture, enabling it to execute CUDA operations natively. The model uses a Gated DeltaNet mechanism as memory a…
Prime Intellect has introduced Prime Mega MoE, a trainable version of DeepSeek’s fused MoE kernel that accelerates forward pass performance by up to 3.11x. The update enables efficient training of Qwen3 MoE models, reduc…
As AI training faces limitations with public and synthetic data, labs are increasingly purchasing digital archives from defunct startups, turning discarded code and workspace records into valuable training material. This…
FlashMap, developed by Polymathic AI and researchers from Princeton and USC ISI, improves large language model inference speed by generating up to 16 tokens simultaneously in a single pass. The technique achieves a 6.14x…
OpenAI engineers use Codex Ultrafast to accelerate incident investigations, reducing analysis time from hours to minutes. The tool enables rapid security assessments, code refactoring, and improved daily workflows. This …
Tsinghua AI has open-sourced AKV, a cache engineering server built on SGLang that improves LLM agent efficiency by dynamically repositioning and dropping KV cache instead of summarizing conversation history. This approac…
OpenAI has released preliminary guidelines for implementing formal safety cases during the training of frontier AI systems, addressing technical protections and operational procedures. The framework includes protocols fo…
This tutorial demonstrates an end-to-end pipeline for augmenting image, text, and audio data with AugLy and testing model robustness. It sets up reproducible synthetic datasets, handles dependency issues, and explores Au…
The OpenAI Foundation announced a program called Public Data for Health to fund creation of high-quality scientific datasets. It will support efforts like collecting data on failed biotech companies' regulatory filings, …