Eintopf-Qwen3.8-27B

Eintopf-Qwen3.8-27B

Eintopf: a German one-pot stew. Everything goes in.

Qwen3.8-27B with nine LoRAs merged in at full strength — all of them trained on the Qwen3.6 line, none of them trained on this model.

This is a negative result, published because the negative result is the interesting part. The adapters move multiple-choice reasoning and do essentially nothing to behaviour. Use Wichtel-Qwen3.6-27B if you want a model that works.

What it does to reasoning

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

Stacking nine buys nothing over one. 173 correct against 175 — inside noise of each other. A sibling finding in this lab is that delta magnitude does not predict damage; this adds that delta quantity does not predict gain.

What it does to behaviour: nothing

47 operator cases — tool selection from a 10-tool set, egirl's system prompt, tool calls parsed out of raw assistant text. right_tool is the headline; delegate asks whether the model hands coding work to a code agent instead of doing it itself.

Base 27, one adapter 28, nine adapters 26. Those are the same number. Every adapter in the stew is invisible on this benchmark, including the character and tool-use ones that were the whole reason for merging them.

The correction this forces

An earlier version of this card led with "adapters port across pretrains" on the strength of the ARC result. The behavioural eval says that claim was too broad, and the control that shows it is the third row above: the delegation adapter scores 3/3 on Qwen3.6-TIES and 0/10 on Qwen3.8 by itself. That is not the other eight adapters drowning it out — it is the same adapter, alone, on a different base, doing nothing.

So the honest version is narrower:

A task vector trained on Qwen3.6 transfers to Qwen3.8 as a shift in multiple-choice accuracy, and does not transfer as a behavioural policy.

Which is a more useful thing to know than the claim it replaces. Weight-space proximity (cosine 0.9405) predicted the ARC transfer correctly and predicted the behavioural transfer wrongly, so basin analysis is not sufficient grounds to skip retraining an operator adapter on the base you intend to ship.

On delegation cases Eintopf reaches for execute_command (7/10) or glob_files (2/10) — it tries to do the coding work itself, which is precisely the habit the delegation adapter exists to break.

One thing it does hold onto: 47/47 no-hallucination across every configuration. It never invents a tool outside the provided set.

Ingredients

All merged at scale 1.0. Order is irrelevant — deltas sum, and that was verified separately (swapping two adapters moved 2 of 299 tasks, which is float non-associativity).

Every adapter is r=32 α=64 over q,k,v,o,gate,up,down. On this architecture those suffixes match the MLPs on all 64 layers but attention on only 16 — three of every four layers use linear attention under different module names — so most of the merged change lands in the MLPs. That is a plausible mechanism for the split result: the adapters can move what the MLPs encode without touching most of the attention routing.

What was merged, and what was not

1,199 tensors in, 1,199 out. 256 module deltas applied. The 333 vision tensors and 15 mtp.* heads are untouched, carried through from Qwen3.8 unchanged.

Merged by summing B@A × (α/r) per module by hand rather than through merge_and_unload, which has silently dropped all 15 MTP heads on this architecture before. The script verifies the tensor count and that every delta found a home, and refuses to write otherwise.

Where Qwen3.8 and Qwen3.6 actually differ

From a per-tensor comparison against Wichtel (cosine, float64):

The vision towers are effectively identical, so anyone hoping Qwen3.8 brings a better encoder should check before doing the work. The divergence is concentrated in the MLPs, deepest in early layers (0, 13–16, relative deltas 0.76–0.79).

Honest limitations

• Use Wichtel if you want the better model. It wins by 7.4 points on ARC and by 21 cases out of 47 on operator behaviour.
• Nine adapters, one scale, no sweep. 1.0 each was chosen because it is what got tested, not because it was tuned. A lower scale might do better; nobody has looked.
• The operator bench was run against Q8_0 with the adapters applied at f16 by llama.cpp, not against these merged f16 weights. ARC was measured the same way.
• The base is Apache-2.0 and so is this. The adapters were trained on a mix of corpora; see each adapter's own card.