Wichtelchen-Qwen3.5-9B

Wichtelchen: the little Wichtel — same house-imp, smaller boots.

The Wichtel-Qwen3.6-27B recipe rebuilt on Qwen3.5-9B: a local operator model that delegates engineering work to a code agent 10/10 — parity with the 27B — and writes Hemlock at 56.1% on hembench, most of the 27B's 67.6% at a third the parameters. Every adapter was retrained on the 9B itself — LoRAs do not transfer across pretrains as behaviour.

Lineage

Qwen/Qwen3.5-9B-Base
  └─ huihui-ai/Huihui-Qwen3.5-9B-abliterated
      └─ nbeerbower/Huihui-Qwen3.5-9B-abliterated-TIES     (+ Grimoire SFT/ORPO, 3-way TIES)
          └─ + Schierling  (Hemlock SFT, Hemlock-SFT-combined)  = Hemlock-Qwen3.5-9B
              └─ + Bubba       (GreatFirewall ORPO — states facts it otherwise sanitises)
                  └─ + delegation  (hands engineering work to a code agent)
                      └─ + egirl       (delegation + Hemlock tool use, ORPO)   ← Wichtelchen

Results — the full ladder, every rung benched

	baseline	+Schierling	+Bubba	+delegation	Wichtelchen
hembench (weighted)	28.5%	57.1%	49.4%	50.6%	56.1%
egirl 47-case	41/47	37	36	36	37/47
— delegate axis	6/10	7	7	9	10/10
censorship (best-of-5)	28.8/29	29.0	28.8	28.6	29.0/29
ARC (deterministic)	61.54	62.54	62.54	62.54	61.87
wiki perplexity	6.656	6.782	6.739	6.717	6.720

The Bubba rung cost 7.7 hembench points at rung 2; the egirl rung's Hemlock preference pairs recovered nearly all of it (49.4 → 56.1) — the same interaction the 27B showed. General capability never moved: ARC stayed in a one-point band across all four rungs and final perplexity is within 1% of baseline.

Limitations

• Call formatting wobble (9B-specific). On 3 of 47 bench cases the model emits a malformed tool call — arguments as a bare string, or an argument hoisted to the top level. The 27B never did this. A lenient parser (or one repair round of preference training) covers it.
• Restraint is the weak behavioural axis (7/10): it sometimes reaches for git_status or execute_command on prompts that want a direct answer. Delegation itself never misfires.
• stdlib is the weak Hemlock axis, as in every model of this family: exact function names are memorised API knowledge. Check calls against the docs.
• Single-turn tool use, and delegation is bound to egirl's tool schema.

MTP head

The 15 mtp.* tensors are restored after every PEFT merge (merge_and_unload() silently drops them and current llama.cpp then refuses the model). 775 tensors — 426 language, 333 vision, 15 MTP, 1 lm_head — verified at every rung.

llama-server -m Wichtelchen-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 4

Related

• schneewolflabs/Wichtel-Qwen3.6-27B — the big sibling
• hemlang/Hemlock-Qwen3.5-9B — this model before the behaviour rungs
• hemlang/Schierling-Qwen3.5-9B-LoRA — the Hemlock adapter
• egirl — the agent this serves