Rnj-1 Instruct

Rnj-1 Instruct is an 8B dense language model built from the ground up by Essential AI with a sharp focus on programming, mathematical reasoning, and scientific problem-solving. Its compact architecture is engineered to punch well above its weight class, delivering elite agentic coding performance that rivals models ten times its size on real-world software engineering benchmarks. The model natively handles function calling and structured tool orchestration through Hermes-format parsing, making it especially well-suited for autonomous coding agents, multi-step technical workflows, and iterative software development pipelines.

The model was released under an Apache 2.0 license with deliberately limited post-training, preserving flexibility for community fine-tuning and domain specialization. On SWE-bench Verified, it scores 20.8%—a figure that places it ahead of much larger competitors on practical code resolution tasks. Its 62.2% score on the BFCL benchmark reflects strong tool-use and function-calling capabilities, while competition math performance on AIME demonstrates solid STEM reasoning. The combination of open weights, compact efficiency, and built-in extensibility makes Rnj-1 Instruct a practical foundation for teams looking to deploy or customize an elite coding model without relying on opaque, heavily post-trained systems.