---
base_model: AutoDecompiler/AutoDecompiler-30B-pscode
pipeline_tag: text-generation
tags:
  - gguf
  - llama.cpp
  - decompilation
  - reverse-engineering
  - pcode
  - qwen3_moe
---

# AutoDecompiler-30B-pscode GGUF

This repository contains an unquantized BF16 GGUF conversion of
[`AutoDecompiler/AutoDecompiler-30B-pscode`](https://huggingface.co/AutoDecompiler/AutoDecompiler-30B-pscode).
It is a format conversion for llama.cpp; the model weights were not fine-tuned
or otherwise modified here.

The model is specialized for turning decompiler P-code/pseudocode into a
high-level source-code draft. It is not intended to decompile raw assembly
directly.

## File

| File | Format | Size | SHA-256 |
| --- | --- | ---: | --- |
| `AutoDecompiler-30B-pscode-BF16.gguf` | GGUF v3, BF16 | 61,095,804,640 bytes (56.89 GiB) | `83d5a42e828e391aac16e68ac2fe7332d7ed4307cf5cf470cb3c33be35c92367` |

The GGUF contains the model's Qwen3 MoE architecture and chat template. Its
metadata advertises a 262,144-token context, but practical context size is
limited by available memory. The weights alone require roughly 57 GiB, with
additional memory needed for the KV cache and runtime workspace.

## Download

```bash
hf download guu3/AutoDecompiler-30B-pscode-GGUF \
  AutoDecompiler-30B-pscode-BF16.gguf \
  --local-dir .
```

## Run with llama.cpp

Use a recent llama.cpp build with Qwen3 MoE and BF16 support:

```bash
llama-server \
  --model AutoDecompiler-30B-pscode-BF16.gguf \
  --alias autodecompiler-30b-pscode-bf16 \
  --host 127.0.0.1 \
  --port 8080 \
  --ctx-size 32768 \
  --n-gpu-layers all \
  --flash-attn on
```

Reduce `--n-gpu-layers` or `--ctx-size` if the model does not fit available
GPU or unified memory.

Send P-code through the OpenAI-compatible endpoint:

```bash
curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "autodecompiler-30b-pscode-bf16",
    "messages": [
      {
        "role": "system",
        "content": "You are a decompilation specialist. Convert the supplied P-code pseudocode into a faithful high-level source representation."
      },
      {
        "role": "user",
        "content": "Recover high-level source code from this P-code pseudocode. Return code only.\n\n<PASTE PCODE HERE>"
      }
    ],
    "temperature": 0,
    "max_tokens": 4096,
    "stream": false
  }'
```

For reproducible evaluation, start with greedy decoding (`temperature: 0`).
Increase `max_tokens` for larger functions, while keeping in mind that very
large functions may be truncated or become impractically slow.

## Conversion provenance

- Upstream model revision:
  `43c73bfabe24c284c31ccb76fc5dc90e5736b5dc`
- Conversion tool: llama.cpp `convert_hf_to_gguf.py`
- llama.cpp revision:
  `48d22e295e2b86b47366c16390794f3e05ba970a`
- Output type: BF16, with tensors that llama.cpp keeps in F32 left in F32
- Tested with llama.cpp build 10360

The conversion is equivalent to:

```bash
python convert_hf_to_gguf.py /path/to/AutoDecompiler-30B-pscode \
  --outfile AutoDecompiler-30B-pscode-BF16.gguf \
  --outtype bf16
```

## Limitations

Treat generated code as an untrusted first-pass draft. In local experiments,
the model recovered useful structure from P-code, but it could emit invalid
identifiers or types and could fail to finish very large functions. Validate
the result against the original binary, compiler diagnostics, and control flow.

See the
[`AutoDecompiler` paper](https://arxiv.org/abs/2606.16162) and the
[upstream model repository](https://huggingface.co/AutoDecompiler/AutoDecompiler-30B-pscode)
for the model and research context.

## License

The upstream model repository did not declare a license at the time of this
conversion. This repository does not assert a new license over the model
weights; consult the upstream authors before redistribution or commercial use.