Model set
Each source was extracted against the base and folded in as its own rotation
block, in priority order. Rescale is the magnitude parity factor applied before
the fold — source magnitudes span nearly three orders of magnitude, and applied
raw the largest of them destroys the model.
Reasoning cost
GSM8K, thinking enabled, 4096-token budget, generated tokens counted directly:
Table with columns: accuracy, median tokens, mean, max, budget overruns | accuracy | median tokens | mean | max | budget overruns |
|---|
| Isometry Terse | 0.85 | 62 | 86 | 329 | 0 / 20 |
| Qwen3.6-27B base | 0.75 | 1168 | 1364 | 4096 | 2 / 20 |
The distributions barely overlap — the merge's worst case runs shorter than a
quarter of base's median.
What it is
The body is composed with an isometric merge. Each source delta is reduced to
the rotation it induces on the base weight inside the subspace it actually
moved, and applied as W → W + Z(R − I)ZᵀW. Because R is orthogonal the
merged weight keeps the norm it started with, so there is nothing to renormalize
afterward and nothing for a low-bit quantizer to spend dynamic range on.
Sources fold as separate rotation blocks in priority order rather than averaging
together, each residualized against the blocks already placed. The AEON
overlay is applied after the body, restricted to output projections.
Same operator as the
Gemma-4 Isometry
models.
Usage
Standard Qwen3.6 chat template, thinking on or off. The efficiency shows with
thinking enabled.
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration
model_id = "maldv/Qwen3.6-27B-Isometry-Terse"
tok = AutoTokenizer.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
model_id, dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "..."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True,
return_tensors="pt", enable_thinking=True).cuda()
out = model.generate(ids, max_new_tokens=1024)
Vision tower and MTP head carry through from base untouched.
Citation
@misc{maldv2026isometryterse,
title = {Qwen3.6 27B Isometry Terse},
author = {Praxis Maldevide},
year = {2026},
note = {Isometric, norm-preserving merge of Qwen/Qwen3.6-27B with shorter chain-of-thought},
howpublished = {\url{https://huggingface.co/maldv/Qwen3.6-27B-Isometry-Terse}}
}
