14B (dense)
Apache 2.0
Competitive coding-focused model post-trained from Qwen3-14B using reinforcement learning, positioned for olympiad-style and verifiable programming tasks. Official release materials report 67.87% Pass@1 on LiveCodeBench v6, a +7.08 point gain over the stated Qwen3-14B baseline (60.79%). Training is described as using 24k verifiable coding problems on 48 B200 GPUs over approximately four days, with a fully released stack that includes model weights, open RL environment, evaluation harness, and experiment logs. The release details iterative context-length training behavior (including extension from 32k to 40k context during training) and evaluation references at longer windows, plus a pipelined verification setup and parallelization notes intended to support reproducibility. Distribution is provided as open-weight text-generation model artifacts on Hugging Face under Apache 2.0, with accompanying technical write-up and code resources in the official Nous release channels.