PrismML Bets on Compact LLMs to Reshape AI Deployment

NVIDIA's Vera Rubin NVL72 system achieved leading performance in its MLPerf Inference v6.1 debut, delivering up to 3.7x higher throughput than the GB300 NVL72 on Qwen3-VL and up to 2.5x on DeepSeek-R1. The results demonstrate the effectiveness of full-stack hardware and software codesign, including enhanced Tensor Cores, NVFP4 precision, and disaggregated serving techniques. A 288-GPU submission across four GB300 NVL72 racks achieved 99% scaling efficiency, and software optimizations alone delivered up to 1.6x performance gains from v6.0 to v6.1.

Researchers at Nature Machine Intelligence have demonstrated a dual-model architecture for autonomous crystal materials research using two lightweight large language models working collaboratively. The approach combines reasoning and scientific tool execution while maintaining computational efficiency and local deployability. The method achieves competitive performance without requiring expensive infrastructure, making advanced materials research more accessible.

Perplexity is deploying OpenAI's GPT-6 Astra model to handle end-to-end system operations, including writing communications, modifying software, and monitoring production infrastructure. The company reports significantly reduced oversight requirements compared to earlier models. This represents a shift toward autonomous AI management of critical business systems.

Alibaba released Qwen3.8-2.4T-A95B as open weights on August 12, 2026, marking the first time a Qwen-Max-class model became publicly available. The 2.4 trillion parameter model uses a hybrid linear-plus-full-attention architecture with 95 billion activated parameters per token and supports up to 262K native context tokens, extensible to 1M. AWS published a deployment guide showing how to run the model on SageMaker HyperPod using vLLM on ml.p6-b300 instances with NVIDIA B300 Blackwell Ultra GPUs.