---
base_model: Qwen/Qwen3.5-4B
license: apache-2.0
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- flutter
- dart
- mobile-development
- gguf
- llama.cpp
---

# RTL-Flutter 0.2

`rtl-flutter-0.2` is a standalone, merged GGUF model for Dart, Flutter, and
mobile-engineering assistance. The LoRA adapter has already been merged into
the Qwen3.5-4B base; users do not need to download or pass a separate adapter.

## Run with llama.cpp

Download this repository and use the GGUF file directly:

```bash
hf download RahnTechLabs/rtl-flutter-0.2 \
  rtl-flutter-0.2.gguf --local-dir ./rtl-flutter-0.2

llama-cli \
  -m ./rtl-flutter-0.2/rtl-flutter-0.2.gguf \
  --jinja \
  --reasoning-budget 0 \
  -p "Explain how Flutter Widget.canUpdate works."
```

For a local OpenAI-compatible server:

```bash
llama-server \
  -m ./rtl-flutter-0.2/rtl-flutter-0.2.gguf \
  --jinja \
  --reasoning-budget 0
```

The file is Q4_K_M quantized and is approximately 2.7 GB. A llama.cpp build
with Qwen3.5 support is required; current llama.cpp releases provide this
architecture.

## Intended use and limitations

This is an experimental domain model for engineering assistance, code review,
debugging explanations, and architecture discussions involving Flutter, Dart,
Android, and iOS. It can produce confident errors, especially on version-
specific APIs and edge cases. Verify answers against the current SDK and
official documentation before shipping production code.

The held-out benchmark and training data are not included in this repository.
Do not put secrets, proprietary code, or personal data into prompts.

## Training and provenance

- Base: `Qwen/Qwen3.5-4B`
- 475 training examples and 25 validation examples
- 2 epochs, learning rate `5e-6`
- LoRA rank 8, alpha 16, dropout 0.05
- bfloat16 training in the project ROCm/PyTorch workflow
- Sources included authorized local Dart/Flutter material and current official
  Flutter and `flutter_bloc` documentation

The published file was produced by merging the project LoRA adapter into a
compatible Q4_K_M base, then requantizing the merged weights to Q4_K_M for
standalone distribution. Requantization can cause a small quality change from
the unquantized merged intermediate.

## License

The base model is distributed under Apache-2.0. This release contains derived
weights, so review the base model terms and ensure that you have the necessary
rights for any local source material before redistributing or deploying it.