Rnj-1 beats 30B models on Kotlin benchmark

Small, specialized models are punching above their weight class — RNJ-1's top-3 finish against 30B+ models signals a meaningful efficiency shift for local Kotlin coding.

• –EssentialAI RNJ-1 hits 75%/81% pass@1/pass@3 at just 8.8 GB, well within single-GPU consumer setups (16 GB VRAM)
• –GPT-OSS 20B still dominates at 85%/95% despite a 12 GB footprint, cementing its position as the local coding model to beat
• –Qwen3.5-35B jumped 18 points in seven months, showing rapid iteration in the open-weights coding space
• –JetBrains Kotlin HumanEval uses 161 expert-written problems verified by Kotlin specialists — a more signal-rich eval than generic HumanEval ports
• –For developers targeting Kotlin specifically, the results suggest capable local inference no longer requires high-end multi-GPU hardware