Today for AI
HOT RADAR
llmHEAT 16.4°

AI CURATED HOT STORY · 10/20/2026

RL Post-trained Qwen3.8-27B Halves Hallucinations and Beats GPT-6 Luna

1 reports archived1 independent sourcesupdated 10/20/2026, 22:08:25
Synthesis & Latest Updates
1 Sources Cross-Validated

An RL post-trained version of Qwen3.8-27B trained on open-source environments significantly improves agent performance, reducing hallucination rates on AA-Omniscience from 30.4% to 15.1% and outperforming GPT-6 Luna-medium on Terminal-Bench. The model features an MLA architecture that fits on a single RTX 4090 with 3.5x more context than the base model, while generating fewer output tokens per task, indicating higher 'intelligence density'.

LATEST/An RL post-trained version of Qwen3.8-27B trained on open-source environments significantly improves agent performance, reducing hallucination rates on AA-Omniscience from 30.4% to 15.1% and outperforming GPT-6 Luna-medium on Terminal-Bench. The model features an MLA architecture that fits on a single RTX 4090 with 3.5x more context than the base model, while generating fewer output tokens per task, indicating higher 'intelligence density'.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·85 pts
    • The RL post-trained Qwen3.8-27B reduces hallucination rates on AA-Omniscience by nearly half (30.4% to 15.1%).
    • Outperforms Muse Glimmer 30B and GPT-6 Luna (high) on Terminal-Bench, τ³-bench, and BFCL v4.
    • The MLA version compresses memory per token from 64 KiB to 18 KiB, enabling single-card RTX 4090 deployment with 3.5x extended context.
    • Median output tokens per solved task dropped from 54.9K (base) to 23.8K, indicating significantly improved inference efficiency.