Elster-Verstand-Qwen3.6-27B

Elster-Verstand-Qwen3.6-27B

<img src="elster.png" alt="Elster" width="380"/>

Elster with Verstand merged in at scale 0.25.

Elster had a documented defect: on a probe of eight reasoning prompts it entered a degenerate loop on one of them, and no standard benchmark could see it. This fixes that, and makes the model substantially more concise in the process.

It fixes the loop

Asked for a Python longest-common-prefix function, Elster reasons correctly for ~2,000 characters and then emits Hemlock code 79 times without ever closing </think> — 89.6% 8-gram repetition, no answer ever produced.

The loop is gone, and the whole suite completes in 40% of Wichtel's reasoning tokens — more concise than either parent, with nothing failing to terminate. Temperature 0, 12k token cap, matched prompts throughout.

Why 0.25, and why not higher

Verstand is a rescued 20% checkpoint whose effective delta is very large — ‖B·A‖ = 7.81 against Federkiel's 0.287 and Stimme's 2.143. It has a narrow usable band, and the boundary is sharp.

ARC-Challenge (299 tasks, paired McNemar against the base):

Reasoning style markers, same prompts:

These two tables describe the same phenomenon. At 0.25 the adapter corrects the base's own reasoning pattern: the structured scaffolding is preserved (70 → 79) and hedging drops (10 → 3). At 0.50 it replaces it — structure collapses to 1, traces flip into distilled DeepSeek-R1 register ("Okay, I need to write a Python function..."), and reasoning quality declines with it.

It is not a gradient. It is a phase change between 0.25 and 0.50, and the ARC decline tracks it. Do not raise this scale. At 1.0 it costs 4 points of ARC — the only statistically significant ARC regression measured across ~24 configurations in this line of work.

The agent benchmarks moved down

This is the cost side, and it is not clean.

Paired McNemar on hembench says none of it is significant — Wichtel → Elster is 3 lost / 2 won (p = 1.000), Elster → Elster-Verstand is 3 lost / 1 won (p = 0.625), and the cumulative Wichtel → Elster-Verstand is 4 lost / 1 won (p = 0.375). Much of the raw drop is task churn rather than loss: Elster gained L3-H-02 and L4-M-01 while losing others.

But five measurements across two independent benchmarks all point the same way, and none of them individually can rule out a real regression. The honest statement is that 47 binary cases and 38 tasks cannot resolve a ~5% agent regression, so this is unmeasured rather than disproven.

Practical guidance: prefer Wichtel for agent and tool-use work. Use this model where thinking mode matters — it is the only one in the line with no degenerate loop and it reasons in 40% of the tokens.

What is not claimed

• Prose is unchanged. Elster's card is explicit that no configuration measurably improved generated prose quality across 240 blind GPT-5 judgements. Nothing here revisits that.
• The probe is 8 prompts. The loop fix is unambiguous — 5681w unclosed becomes 722w with a correct answer, reproducible at temperature 0 — but "0/8 unclosed" is not a guarantee of no loops anywhere.
• Verstand is a 20% checkpoint on a schedule that never annealed, rescued from a run killed by a memory guard. A finished run would likely want a different scale entirely.

Merging

Merged by hand rather than with peft's merge_and_unload(), which silently drops the 15 mtp.* tensors and produces a model llama.cpp will not load. Verified at 1199 tensors (850 language, 333 vision, 15 MTP, 1 lm_head) before writing. Stacking order is irrelevant — LoRA deltas are a sum.

Full lineage: Qwen3.6-27B → TIES → Vorarbeiter → +Schierling (Hemlock) → +egirl = Wichtel → +Stimme 0.50 +Federkiel 1.00 +Chud 1.00 = Elster → +Verstand 0.25 = this.