Building a state-of-the-art 32B reasoning model on AMD Instinct™ MI325X GPUs using only synthetic data. SAND-MathScience-DeepSeek-Qwen32B achieves 78.33% on AIME25, surpassing Qwen3-32B, with fully open-sourced model weights, datasets, and training code.

AMD provides Day-0 support for slime, the SGLang-native RL framework verified for large-scale MoE models up to 355B parameters. Features APRIL optimization achieving 40% throughput improvement, Docker deployment, and multi-turn kernel-agent training on AMD Instinct™ GPUs.

Introducing Instella-Math — AMD's first fully open reasoning language model, trained end-to-end with long chain-of-thought reinforcement learning on 32 MI300X GPUs. Optimized for complex reasoning and math, with all code, weights, and datasets released for the community.

Introducing GEAK, AMD’s AI-driven Triton kernel generation framework and evaluation benchmarks. GEAK leverages frontier LLMs with inference-time scaling to automatically produce efficient, accurate GPU kernels for AMD Instinct™ GPUs, achieving up to 63% correctness and 2.59× speedups over reference implementations.

Introducing 1D binary image latents for text-to-image generation, achieving up to 32× token reduction and competitive 1024×1024 performance with faster, scalable training and inference.

Introducing Instella-Long, a fully open 3B-parameter language model supporting 128K context length, trained on 64 Instinct MI300X GPUs with open-sourced weights, data, and code, showcasing AMD hardware’s scalability and enabling community-driven long-context AI research.

Overview of Volcano Engine Reinforcement Learning (verl) for large-scale RLHF on AMD Instinct GPUs. Learn about AMD-specific optimizations, Docker setup, and training scripts for single-node and multi-node deployments, plus performance benchmarks on MI300X.

Introducing Instella-VL-1B, the first open-source vision-language model trained on MI300X GPUs, combining a CLIP-based vision encoder with a 1B-parameter language model and outperforming comparable open-weight models on general and OCR benchmarks, with full weights, data, and code released for the community.

Introducing Instella, a fully open family of 3B-parameter language models trained from scratch on MI300X GPUs, outperforming similar open models and achieving competitive performance with state-of-the-art open-weight models, with all weights, data, and code released to the community.

Introducing AMD OLMo, a fully open series of 1B-parameter language models trained from scratch on MI250 GPUs with 1.3T tokens, featuring pre-trained, SFT, and DPO checkpoints, empowering the AI community with reproducible training details and efficient deployment on AMD hardware.