What changed
Only these weights differ from the base model, 2,048 numbers in all:
Table with columns: Layer, Neuron, Change| Layer | Neuron | Change |
|---|
| 14 | 2172 | output column mlp.down_proj.weight[:, 2172] ×−20 |
| 15 | 1572 | output column mlp.down_proj.weight[:, 1572] rewritten (the change is 10.4× the original column's length) |
The exact new values are in edits/.
Results
Held-out terse score 0.26 → 0.355. Each answer scores 0 if it is wrong or cut off, otherwise
how far its length moved from the model's normal answer toward its "answer briefly" answer; the
post describes the score in full. Thinking off.
Side effects we measured, against the unedited model on the same questions: IFEval prompt-level
strict (thinking off) 53.8% → 48.2% (−5.5 ± 2.2), median answer 122 → 91 words, losing mostly
instructions that ask for extra text, such as postscripts, sections and minimum lengths; MMLU-Pro with thinking
on 32.9% → 31.4% (−1.4 ± 1.9).
Usage
Load it like the base model:
from transformers import AutoModelForImageTextToText, AutoTokenizer model = AutoModelForImageTextToText.from_pretrained("rilixai/Qwen3.5-0.8B-Beaker-terse-3230a4ea", torch_dtype="bfloat16")tokenizer = AutoTokenizer.from_pretrained("rilixai/Qwen3.5-0.8B-Beaker-terse-3230a4ea")messages = [{"role": "user", "content": "What is 12 × 7?"}]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
or serve it with vLLM: vllm serve rilixai/Qwen3.5-0.8B-Beaker-terse-3230a4ea.
License
Apache-2.0, as the base model. This is a modified version of
Qwen/Qwen3.5-0.8B: only the weights listed above
were changed. See LICENSE.