RL-post-trained Qwen3.8-27B reportedly cuts its AA-Omniscience hallucination rate from 30.4% to 15.1%

A user says reinforcement-learning post-training of Qwen3.8-27B with open-source environments cut its hallucination rate on AA-Omniscience from 30.4% to 15.1%. They say a compressed copy of its multi-head latent attention (MLA) version fits their 24 GB RTX 4090 and holds about 3.5x more context than base Qwen. They also report it tops GPT-6 Luna (high) on several agent benchmarks, while Luna (high) still leads on DeepSWE.