NVIDIA

NVIDIA — Hardware

Changelogs

NVIDIA Launches Free GPU-Accelerated Endpoints for Alibaba's Qwen3.5 397B Vision-Language Model//Alibaba's Qwen3.5 is now available on NVIDIA's free GPU-accelerated endpoints and through API integration, bringing a 397B-parameter multimodal model with UI navigation capabilities to developers. The release includes support for containerized deployment via NVIDIA NIM and fine-tuning tools through the NeMo framework for domain-specific customization.

featuremodelapiplatformintegration

NVIDIA Achieves 1.6x Training Throughput Gains with NVFP4 Low-Precision Format on Blackwell GPUs//NVIDIA has published results demonstrating that NVFP4, a 4-bit low-precision training format, delivers up to 1.6x higher throughput and substantial memory savings while maintaining accuracy comparable to BF16 on large-scale model training. The results come from experiments training Llama 3 8B and an internal research model on 1 trillion tokens using the open-source NeMo Megatron Bridge framework on B200 GPUs.

featureperformancemodelsdkopen-source

NVIDIA TensorRT LLM adds AutoDeploy beta, automating model-to-inference optimization pipeline//NVIDIA announced AutoDeploy, a new beta feature in TensorRT LLM that automatically compiles PyTorch models into inference-optimized graphs without manual reimplementation. The compiler-driven approach enables deployment of new architectures at launch while supporting over 100 text-to-text LLMs, VLMs, and state space models.

featuresdkapiperformance

Kimi K2.5 multimodal VLM now available on NVIDIA GPU-accelerated endpoints with free prototyping access//NVIDIA has made Kimi K2.5, a 1-trillion parameter multimodal vision language model, available for free prototyping on its GPU-accelerated endpoints via build.nvidia.com. The model supports text, image, and video inputs and can be deployed via vLLM or fine-tuned using the NVIDIA NeMo Framework for production use cases.

featuremodelapiintegrationplatform