RL post-trained Qwen3 27B halves hallucinations, runs on a single RTX 4090
rohanpaul_ai · x · 2026-10-20
An RL post-trained Qwen3.8-27B built with open-source environments posts striking results: hallucination rate on AA-Omniscience drops from 30.4% to 15.1%, it beats GPT 6 Luna-medium on DeepSWE, and scores higher on Terminal-Bench 2.1 while writing a median 23.8K output tokens per solved task (vs 54.9K for base Qwen). The author, who tried an unreleased version, notes a much higher "intelligent density" with far less output.
- Tops Muse Glimmer 30B and GPT-6 Luna (high) on Terminal-Bench, τ³-bench, ClawEval and BFCL v4; Luna still leads DeepSWE at 53.1%
- An MLA version converts 16 full-attention layers to multi-head latent attention, cutting per-token memory from 64 KiB to 18 KiB; the compressed build fits a single 24GB RTX 4090 with 3.5x more context than base
More from Models
- User claims GPT-6 Astra beats Opus 5.5 on frontend coding, unverified — Angaisb_ · 2026-10-23
- LLMs weren't built for code: one architecture serving prose and strict logic degrades both — gerardsans · 2026-10-23
- User claims he got an AI to admit its parameters 'exclude altering the status quo' — lordcycy · 2026-10-23
- Perplexity ships Decider for Mac, running its 4B decision model locally with no API key — perplexity_ai · 2026-10-23
- Open-source Bonsai 2 9B gets 471% faster on Macs after one-week community optimization challenge — gajesh · 2026-10-23
- SuperGrok in Cursor: user burns through weekly quota, 3-day wait for reset — mhdfaran · 2026-10-23