Rnj-1 is a family of 8B parameter open-weight, dense models trained from scratch by Essential AI, optimized for code and STEM with capabilities on par with state-of-the-art open-weight models. The base model rnj-1 and its instruction-tuned variant rnj-1-instruct perform well across a range of programming languages and boast strong agentic capabilities within frameworks like mini-SWE-agent, while also excelling at tool-calling. The models additionally exhibit strong capabilities in math and science. The models demonstrate strong code generation abilities across tasks like HumanEval+, MBPP+, BigCodeBench, and LiveCodeBench v6, competing with the strongest open weight models and sometimes outperforming even larger models. Rnj-1-instruct particularly dominates on agentic coding tasks, scoring 20.8 percent on SWE-bench Verified in bash-only mode, which exceeds Gemini 2.0 flash and Qwen2.5-Coder 32B under the same framework. The model is also able to use profilers to iteratively improve code performance and surpasses comparable models in tool use performance. Both models are designed to be extended and specialized by the community, with limited post-training to allow for further domain customization. The base model uses global attention and YaRN for long-context extension, was pre-trained on 8.4 trillion tokens with an 8K context length, and can be extrapolated to support up to 128K context through configuration modifications.