Hemlock-Qwen3.5-9B

Hemlock-Qwen3.5-9B

A 9B model that writes Hemlock — the 9B counterpart of Hemlock-Qwen3.6-27B.

Schierling-Qwen3.5-9B-LoRA (SFT on hemlang/Hemlock-SFT-combined) merged at 1.0 into Huihui-Qwen3.5-9B-abliterated-TIES, with the 15 mtp.* tensors restored from the base after the merge — peft merge_and_unload() drops them, and current llama.cpp refuses to load the model without them. 775 tensors total (426 language, 333 vision, 15 MTP, 1 lm_head), verified against the base's tensor inventory.

Results — hembench 57.1%

Every generated program is executed by the real interpreter and its stdout compared against expected output (Q8_0, greedy):

For scale: the 27B TIES scored 10.1% before its Hemlock training and 65.0% after. This 9B lands at 57.1% — most of the 27B's Hemlock capability at a third the parameters.

No measurable general-capability cost from the adapter: ARC 61.5→62.5 (n=299), wiki perplexity +1.9%.

Limitations

• stdlib is the weak axis (1/5), same as the 27B family: standard-library work needs exact function names and signatures — memorised API knowledge rather than transferable syntax. Check calls against the docs.
• Tool-use restraint measured slightly below the pre-SFT base (8/10 vs 10/10 on a 47-case bench) — the known cost of direct code-generation SFT on an operator model. Preference training repairs this; it is the next rung of this model's ladder.

Serving

llama-server -m Hemlock-Qwen3.5-9B-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on

The MTP head supports --spec-type draft-mtp speculative decoding.