Qwen3.8-27B 强化学习后训练模型在多项 Agent 基准上超过 GPT-6 Luna high
Some of the most interesting agent results this year are coming out of post-trains of small open models.
Rohan Paul 介绍了一个基于 Qwen3.8-27B 做强化学习后训练的开源环境模型，在 AA-Omniscience 上把幻觉率从 30.4% 降到 15.1%。该模型在 Terminal-Bench、τ³-bench、ClawEval 和 BFCL v4 上超过 Muse Glimmer 30B 与 GPT-6 Luna high，但 DeepSWE 仍由 Luna 以 53.1% 领先；它解出任务的中位输出为 23.8K token，基础 Qwen 为 54.9K。作者还提到 MLA 版本把 16 层全注意力改为多头潜在注意力，每 token 内存占用从 64 KiB 降到 18 KiB，压缩版可跑在单张 24 GB RTX 4090 上，上下文约为基础 Qwen 的 3.5 倍。
正文
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.