Qwen3.8-27B RL Post-Train Cuts Hallucination Rate in Half
The notable result is efficiency: matching frontier models on agent benchmarks while emitting far fewer tokens per task, which directly cuts inference cost for agentic workflows. Developers running local agent stacks can fit this open model on one RTX 4090 and get longer sessions via the MLA version. Teams choosing between open small models and GPT-6 Luna now have benchmark-level evidence that post-trained Qwen variants compete on most agent tasks.
A new RL post-trained Qwen3.8-27B model, built with open-source environments, lowers hallucination on AA-Omniscience from 30.4% to 15.1%. On Terminal-Bench 2.1 it writes a median 23.8K output tokens per solved task versus 54.9K for base Qwen. It beats Muse Glimmer 30B and OpenAI's GPT-6 Luna (high) on Terminal-Bench, τ³-bench, ClawEval and BFCL v4, while GPT-6 Luna (medium) still leads DeepSWE at 53.1%. A released MLA variant converts 16 full-attention layers to multi-head latent attention, shrinking per-token memory from 64 KiB to 18 KiB so the compressed model fits on a single 24 GB RTX 4090 with roughly 3.5x more context than base Qwen.
Some of the most interesting agent results this year are coming out of post-trains of small open models.
A new RL post-train Qwen3.8-27B model trained with an open-source environments cuts the hallucination rate on AA-Omniscience from 30.4% to 15.1%.
the compressed copy of the mla version fits on my single 24 GB RTX 4090 and holds about 3.5x more context than base Qwen.
It beats GPT 6 Luna-medium on DeepSWE, placing in between Luna 6 medium and high. scores higher on Terminal-Bench 2.1 while writing a median 23.8K output tokens per solved task, where base Qwen writes 54.9K.
Tried an unreleased version of this model, and really felt the comparative intelligent density, that it writes a lot less.
tops Muse Glimmer 30B and OpenAI’s GPT-6 Luna (high) on Terminal-Bench, τ³-bench, ClawEval and BFCL v4 (Berkeley Function Calling Leaderboard), though Luna still wins DeepSWE at 53.1%.
They also released an MLA version, which is the same model with its 16 full-attention layers converted to multi-head latent attention, which shrinks the memory each token takes from 64 KiB to 18 KiB so far more or longer sessions fit on one GPU.
