Instella

Instella is a fully open 3-billion-parameter language model family designed to deliver strong general-purpose language understanding and generation. Developed and trained by AMD on Instinct™ MI300X GPUs, it targets researchers, developers, and organizations that require transparent, reproducible, and high-performing models without restrictive licensing. Instella-3B provides a foundation for tasks such as text generation, summarization, code assistance, and conversational AI in both research and production environments.

Instella stands out through its complete openness: AMD releases model weights, training hyperparameters, datasets, and code, enabling full inspection, customization, and extension. The models are trained from scratch using the AMD ROCm software stack and modern efficiency techniques including FlashAttention-2, Torch Compile, and Fully Sharded Data Parallelism (FSDP) with hybrid sharding for scalable training across large GPU clusters. Benchmarks show that Instella-3B significantly outperforms existing fully open models of similar size and delivers competitive performance relative to open-weight models such as Llama-3.2-3B, Gemma-2-2B, and Qwen-2.5-3B, including their instruction-tuned variants.