The Qwen3.8-27B reinforcement-learning post-trained model surpasses GPT-6 Luna high on multiple Agent benchmarks
Rohan Paul introduced an open-source environment model built on Qwen3.8-27B with reinforcement-learning post-training, which reduced the hallucination rate on AA-Omniscience from 30.4% to 15.1%. The model surpasses Muse Glimmer 30B and GPT-6 Luna high on Terminal-Bench, τ³-bench, ClawEval, and BFCL v4, but Luna still leads on DeepSWE at 53.1%; its median output for solved tasks is 23.8K tokens, compared with 54.9K for base Qwen. The author also mentions that the MLA version replaces 16 layers of full attention with multi-head latent attention, reducing memory usage per token from 64 KiB to 18 KiB; the compressed version can run on a single 24 GB RTX 4090, with a context about 3.5 times that of base Qwen.
Why read this
AI score 63
The author compares the performance of small-model RL post-training on Agent tasks based on first-hand testing, and gives specific differences in memory usage and output tokens.
Topics
Body
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.