Qwen3.8-27B-antislop · GGUF
GGUF conversions of Qwen3.8-27B-antislop — the calibrated
expert-rubric GRPO model (step 100). Text-only target; the vision tower and MTP head are excluded (--no-mtp).
| File | Size | Quantization |
|---|---|---|
Qwen3.8-27B-antislop-Q4_K_M.gguf |
15.41 GiB | Q4_K_M |
Qwen3.8-27B-antislop-Q8_0.gguf |
26.63 GiB | Q8_0 |
Converted and quantized with llama.cpp commit efe3626b119ff81033e309c3ba07a770bef2c546. Static quantization
(no importance matrix). File hashes and conversion details are in
weights_manifest.json.
Why this model
It was trained to write less like an LLM and more like a person — reinforcement learning against expert-aligned rubrics (a reimplementation of Meta FAIR's RL-XAR / unslop). On the full EQ-Bench 4 protocol it reaches Elo 1320.1 (95% CI 1309–1332), rank 4 of 29 (measured on the bf16 checkpoint; 4-bit quantization may shift quality). See the model card for the method, the prose benchmark and the trait profile.
Usage (llama.cpp)
llama-cli -m Qwen3.8-27B-antislop-Q4_K_M.gguf -p "Finish this tweet naturally: Monday mornings are" -n 64 --temp 0.8
Thinking is disabled in the intended prose/rewrite use. Keep the source facts and check the result before publishing; a language model can still omit or change a detail.