Qwen3.8-27B post-train cuts AA-Omniscience hallucinations to 15.1% with MLA variant

An unnamed developer releases an RL post-trained Qwen3.8-27B model that cuts the hallucination rate on AA-Omniscience from 30.4% to 15.1%, according to results shared by Rohan Paul on X. The post-train uses open-source environments.

The developer also releases an MLA version, the same model with its 16 full-attention layers converted to multi-head latent attention, which shrinks per-token memory from 64 KiB to 18 KiB.

Rohan Paul says the compressed MLA version fits on a single 24 GB RTX 4090 with about 3.5x more context than base Qwen. He also reports that the model beats GPT-6 Luna-medium on DeepSWE, while Luna still scores higher at 53.1%. On Terminal-Bench 2.1 the model scores higher than base Qwen, with a median of 23.8K output tokens per solved task versus 54.9K for base Qwen.