Tag: #llm-infrastructure13 results
DeepSeek has introduced dynamic pricing linked to electricity costs, highlighting the growing synergy between AI compute demand and power grid load management. This 'compute-power synergy' is becoming a critical factor for AI infrastructure sustainability.
Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model designed for autonomous software engineering. It features a 1-million-token context window and was trained entirely on Chinese chips.
Former OpenAI research scientist Vinces Yao has joined Tencent to lead the AI Infra and Large Language Model departments. He will report to Tencent President Martin Lau and focus on long-term AI development.
Amazon is reportedly working on a new project codenamed 'Moonraker' to transform its Alexa voice assistant into a sophisticated AI agent. Internal documents reveal concerns regarding the project's extremely high development and operational costs.
Tencent has officially released its Hy3 model, marking a significant step in its AI strategy. The model is being integrated into internal products like WorkBuddy and Yuanbao to improve task success rates and reduce hallucinations.
PyTorch Monarch now supports AMD GPUs, enabling single-controller distributed training on ROCm. This update aims to simplify the scaling of large language model training across massive GPU clusters.
Beijing has officially established a new AI factory infrastructure project. The facility aims to achieve 100,000 P-Flops of computing power and produce 10 trillion tokens daily.
quicktok is a new C++ based BPE tokenizer that achieves 4–11× faster performance than tiktoken while remaining byte-identical. It supports major models including Llama-3, Qwen2.5/3, and various GPT-based encodings.
Xiaomi's MiMo team revealed the technical optimizations behind their V2.5 model, including KVCache management and multi-modal inference improvements. These updates allow for permanent API price reductions while maintaining profitability.
DeepSeek has officially announced that the 75% price reduction for its flagship V4-Pro model will be permanent. This move maintains developer access costs at one-quarter of the original pricing structure.