Open-Sourced 27B Agent Model Halves Hallucination Rate, Runs on a Single 4090

A lab open-sources an RL post-trained Qwen3.8-27B built for agents, cutting hallucination rate from 30.4% to 15.1% (below GPT-6), beating GPT 6 on DeepSWE, halving agent token usage, and running on a single RTX 4090.

2026-10-20 ~ 2026-10-21 · 2 related posts

• RL post-trained Qwen3 27B halves hallucinations, runs on a single RTX 4090 — rohanpaul_ai · 2026-10-20
• Lab open-sources 27B agent model: 15.1% hallucination rate, half the tokens of its base — lukas_m_ziegler · 2026-10-21