Downloads · 30 days
14
15% of all-time downloads
DragonBophades/Elster-Vernunft-Qwen3.6-27B
Elster-Vernunft-Qwen3.6-27B is a image-text-to-text model from DragonBophades. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
Elster with the Vernunft reasoning LoRA merged in at full scale.
Downloads · 30 days
14
15% of all-time downloads
All-time downloads
96
Public
Parameters
27.4B
54.7 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors54.7 GB · 100%
From the Hugging Face model README
Elster with the Vernunft reasoning LoRA merged in at full scale.
It thinks 8× less than Elster, in Elster's own voice, and gives up nothing measurable to do it.
The predecessor, Elster-Verstand, merged the original reasoning adapter at 0.25. That cap
was not a stylistic choice — the original adapter was trained on traces that are 76% R1-register
first-person self-talk, and at higher scales it stopped correcting the base's reasoning and
started replacing it:
| scale | ARC (299, paired) | reasoning style |
|---|---|---|
| 0.25 | +0.67 | structured 70 → 79, R1 10 → 3 |
| 0.50 | −1.00 | structured 70 → 1, R1 10 → 27 |
| 1.00 | −4.01 (p = 0.036) | — |
Vernunft is the same 30,096 traces rewritten into the base model's own structured register. The cliff is gone, so the cap has nothing left to protect:
| scale | ARC | vs base | McNemar p |
|---|---|---|---|
| 0.25 | 64.88 | −0.33 | 1.000 |
| 0.50 | 65.22 | 0.00 | 1.000 |
| 0.75 | 64.55 | −0.67 | 0.824 |
| 1.00 | 64.88 | −0.33 | 1.000 |
Base Wichtel scores 65.22. Flat everywhere — and disagreement grows with scale (7 tasks flipped at 0.25, 27 at 1.0), so the adapter is doing real work at full strength. It simply wins as often as it loses.
Eight held-out prompts (code, math, logic, one trivial), temperature 0, matched 12k token cap:
| reasoning | median | R1 self-talk | structured share | unclosed | |
|---|---|---|---|---|---|
| Elster | 6,725 w | 702 w | 55 | 71% | 0/8 |
| Elster-Verstand @0.25 | 3,654 w | 550 w | 33 | 73% | 0/8 |
| this model @1.00 | 803 w | 74 w | 0 | 100% | 0/8 |
Against Elster-Verstand directly: 4.6× shorter reasoning, R1 markers 33 → 0, structured share
73% → 100% — at four times the adapter dose.
Elster-Verstand barely repaired anything. It cut reasoning 1.8× and moved the register two points, because 0.25 was all it could safely take. That is the argument for rewriting the data rather than turning the dial down.
The effect also transfers: Vernunft was trained on Wichtel, and produces near-identical numbers here on Elster (784 w / 100% structured on Wichtel, 803 w / 100% here). The register fix is not base-specific.
Honest accounting, all measured:
| axis | result |
|---|---|
| ARC-Challenge | flat at every scale, p ≥ 0.824 |
| Hemlock coding (38 tasks, executed) | 47.1% → 47.5% — unchanged |
| thinking length | 8.4× shorter |
| R1 register | eliminated |
| agentic tool-calling | worse — see below |
On a 47-case operator bench (tool definitions in the system prompt, calls parsed from raw text, temperature 0), the adapter costs tool-selection accuracy:
| base nothink | +adapter | base thinking | +adapter | |
|---|---|---|---|---|
| right tool | 47/47 | 44/47 | 42/47 | 38/47 |
| args ok | 47/47 | 46/47 | 45/47 | 41/47 |
Do not put this in an agent loop. Vernunft was trained on bigdenker — math and general
reasoning — so tool-calling prompts are out of its training distribution, and it shows. If you
want an operator model, use Elster.
Note also that on those short operator prompts the reasoning got slightly longer (177 → 224 chars), the opposite of the headline result. Both are real: on two-sentence reasoning there is nothing to compress, and the adapter's structural habits add tokens rather than removing them. The 8× reduction applies to genuine reasoning work.
| base | DragonBophades/Elster-Qwen3.6-27B |
| adapter | schneewolflabs/Vernunft-Qwen3.6-27B-LoRA at 1.0 |
| method | merge_and_unload, bf16, on CPU |
| tensors | 1,184 in → 1,184 out, 333 vision preserved |
The adapter covers the MLP projections on all 64 layers but attention q/k/v/o on only 16. That is Qwen3.6's hybrid attention — three of every four layers use linear attention under different module names, which the standard projection targets never match. Most of the work is in the MLPs.
Inherited caveat: Elster ships with 1,184 tensors rather than the architecture's 1,199 — the
15 mtp.* speculative-decoding heads were already absent before this merge. Nothing here removed
them, and nothing here restores them.
Vernunft-SFT is Verstand-SFT's
bigdenker split rewritten by Wichtel-Qwen3.6-27B at temperature 0 — same prompts, same answers,
only the reasoning changed: 398 → 99 words mean, R1 register 76% → 0.3%, structured form 24% →
88.6%. Every row keeps reasoning_original, so the change is auditable.