Developer AlexWortega (rawmodels) has released the open model Qwen3.8-27B-antislop — a post-training of Qwen/Qwen3.8-27B (27B parameters, Apache-2.0) using the GRPO method, so that texts sound "less like an LLM, more like a human." The reward is built on frozen expert rubrics and a "length-blind" judge, and on EQ-Bench 4 the model ranked 4th out of 29 with an Elo of 1320.1. Along with the weights and GGUF quantizations for llama.cpp, the complete RL environment unslop-harbor with 200 tasks, a sha256-fixed judge prompt, and REPRODUCE.md has been published — the process is now reproducible by other teams, not just described in a blog post.
What happened
The model, weights, and environment have been published on Hugging Face. The base was Qwen/Qwen3.8-27B, trained using the GRPO method: a "length-blind" judge based on Claude Opus assigns a quality score from 0 to 10 against frozen expert rubrics, and the final reward is calculated using the formula reward = L(words, target) · σ(a_d·(q − μ_d)). On held-out prose, the calibrated judge gave antislop 8.28/10 compared to 6.31 for the base Qwen3.8-27B and 7.10 for Altworld/Hemmingway-1; in pairwise comparisons, the score was 50–0 and 43–6. Release artifacts include BF16 and mxfp4-a16 weights under Apache-2.0, GGUF quantizations Q4_K_M and Q8_0 for llama.cpp, and a Space where the model can be tested in the browser. Separately, the RL environment rawmodels/unslop-harbor was released: 200 Harbor tasks (100 tweets and 100 HumanEval functions, half of the tasks in Russian), where code must pass tests before being evaluated by the claude-opus-5.5 judge; the judge prompt is fixed by a sha256 hash, and the package includes REPRODUCE.md, scores, and the texts themselves.
Context
The problem addressed by the release is shown in research from Meta FAIR: standard LLM judges in pairwise comparisons prefer model text to human text in 63.5–84.6% of cases, and with standard rubrics — up to 100%. This implies that standard RLHF systematically amplifies "slop": the model is optimized for the judge's preferences rather than human text, and this is a structural evaluation problem, not a defect of a specific model. The RL-XAR method from the same work reorients the reward: expert rubrics are calibrated on human texts against AI texts, then frozen and serve as a quality standard. AlexWortega's release is an independent implementation of RL-XAR on an open model via GRPO; this level of openness for RL post-training is rare, because teams usually publish weights without the environment, reward, and judge prompt.
Why this matters for the industry
Product teams are interested not so much in Elo as in the package of three things: open 27B weights, a public RL environment with a reproducible reward, and a ready-made Space. The recipe for "human-like" text has become public, which reduces the cost of vertical post-training and local distribution via GGUF and llama.cpp for startups. At the same time, the moat for those who defended themselves only with the claim "we don't write like robots" is eliminated: the method is now a commodity, and defense must be provided by one's own expert rubrics for one's niche. Forks of RL-XAR-like rewards on other open models and in other domains are expected to appear — Russian-language text is already included in Harbor, with half of its tasks in Russian. If independent replications confirm the scores, rubric-calibrated rewards may enter standard post-training alongside preference methods, and "non-slopiness" will become a standard evaluation metric.
Why this matters for users
The release can be checked today, and it is pure self-hosting. First, it is worth running your own texts through the Space on Hugging Face and comparing the result with your current stack; then download the GGUF quantization Q4_K_M (15.41 GiB) or Q8_0 (26.63 GiB) and set up inference on llama.cpp as a local backend for writing or an editor plugin (release weights are distributed under Apache-2.0). The practical effect is shorter and less template-like text without API costs, if the generations of current tools "sound like a neural network." For those doing their own RL experiments, unslop-harbor will be useful as a public reward: the reward formula, judge prompt, and REPRODUCE.md allow building an experiment according to instructions, not from scratch.
What is currently unknown / limitations
Superiority over frontrunners is not statistically established: the 95% confidence interval on EQ-Bench 4 (1309–1332) overlaps with the results of GPT-5.5 (1314.4) and Claude Opus 4.7/4.8 (1307.9 and 1277.8), so statements like "outperformed GPT-5.5 and Claude Opus" are not justified. EQ-Bench 4 is a stylistic metric, and one measurement cannot extrapolate a general advantage over frontrunners; there are no independent replications of numbers like 8.28/10 vs 6.31 yet. Finally, the release describes only the self-hosting scenario: hosted API, pricing, and latency measurements are absent from the source materials.
Sources
- AlexWortega/Qwen3.8-27B-antislop — Hugging Face (model card)
- rawmodels/unslop-harbor — Hugging Face (RL environment, 200 Harbor tasks)
- Unslopping AI — RAM @ Meta AI (original RL-XAR method)
Author
Look at AI, editorial team