---
license: apache-2.0
base_model:
- Ailiance-fr/SchGen-Qwen3.6-27B-EU
- Qwen/Qwen3.6-27B
datasets:
- microsoft/SchGen_dataset
language:
- en
library_name: gguf
tags:
- pcb
- eda
- kicad
- schematic-generation
- schgen
- gguf
- qwen3
- llama.cpp
- ollama
---

# SchGen-Qwen3.6-27B-EU — GGUF

GGUF quantizations of
[`Ailiance-fr/SchGen-Qwen3.6-27B-EU`](https://huggingface.co/Ailiance-fr/SchGen-Qwen3.6-27B-EU)
— a sovereign EU KiCad schematic-generation model (QLoRA fine-tune of
`Qwen/Qwen3.6-27B`, SchGen method, [Luo et al., 2026](https://arxiv.org/abs/2605.30345)) —
for **llama.cpp** and **Ollama**.

## Files

| File | Quant | Size |
|------|-------|------|
| `SchGen-Qwen3.6-27B-EU-Q4_K_M.gguf` | Q4_K_M | ~16.5 GB |
| `SchGen-Qwen3.6-27B-EU-Q8_0.gguf` | Q8_0 | ~28 GB |

## ⚠️ Runtime requirement (read this)

The base architecture is **`qwen3_5`** (Qwen3.6, *hybrid linear/state-space
attention*) — a **recent** arch. You need a **recent llama.cpp / Ollama** that
registers `LLM_ARCH_QWEN35`. **Verified: Ollama 0.19.0** loads and runs it.
Older runtimes fail at load with *"unknown model architecture"*. On CPU the 27B
is slow; use a GPU for usable throughput.

## Usage — Ollama

```bash
# pull the file, then:
cat > Modelfile <<EOF
FROM ./SchGen-Qwen3.6-27B-EU-Q4_K_M.gguf
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0
EOF
ollama create schgen -f Modelfile
ollama run schgen "Generate a KiCad schematic for an RC low-pass filter, 1kHz cutoff."
```

## Usage — llama.cpp

```bash
llama-cli -m SchGen-Qwen3.6-27B-EU-Q4_K_M.gguf \
  -p "Generate a KiCad schematic with an LM358 inverting amplifier, gain -10." \
  -n 2048 --temp 0
```

The model emits **executable Python in a schematic DSL** (4 primitives:
`add_schematic_symbol`, `get_pin_location`, `add_label`, `connect_pins` +
`write_out_all_wires()`). Run with thinking disabled. Output **must** be checked
with KiCad ERC/DRC — it is an assistant, not autonomous. See the
[main model card](https://huggingface.co/Ailiance-fr/SchGen-Qwen3.6-27B-EU) for
evaluation, limitations and citation.

## License

Apache-2.0 (see `NOTICE`). Derivative of `Qwen/Qwen3.6-27B` (Apache-2.0),
trained on `microsoft/SchGen_dataset` (MIT); method = SchGen (MIT).