---
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.5-4B/blob/main/LICENSE
pipeline_tag: image-text-to-text
base_model:
- Qwen/Qwen3.5-4B
tags:
- qwen
- qwen3.5
- vision-language
- image-text-to-text
- custom
- mosslight
---

# Mosslight 4B

Mosslight 4B is a fine-tuned, merged derivative of Qwen3.5-4B, packaged in
Hugging Face Transformers format for local inference, serving, and downstream
experimentation.

This repository contains the model weights, tokenizer, chat template, and
multimodal preprocessor files needed to load the model with compatible Qwen3.5
tooling.

## Model Details

- **Model name:** Mosslight 4B
- **Model ID:** `ttrpg/mosslight-4b`
- **Base model:** `Qwen/Qwen3.5-4B`
- **Derivative type:** fine-tuned and merged full-weight release
- **Architecture:** `Qwen3_5ForConditionalGeneration`
- **Model type:** vision-language causal generation
- **Parameters:** approximately 4B
- **Native context length:** 262,144 tokens, as inherited from the base config
- **License:** Apache 2.0, inherited from the base model

## Lineage

This model is a fine-tuned, merged derivative of Qwen3.5-4B from Alibaba
Cloud/Qwen. The original Apache 2.0 license is preserved in `LICENSE`, and
derivative attribution is documented in `NOTICE`.

Training and merge details should be completed before publishing a final public
version.

## Training Details

- **Base checkpoint:** `Qwen/Qwen3.5-4B`
- **Fine-tuning method:** TODO
- **Training data:** TODO
- **Merge method:** TODO
- **Output format:** merged full weights in sharded Safetensors format
- **Post-training evaluation:** TODO

## Files

- `config.json`: model architecture and multimodal configuration.
- `model.safetensors-00001-of-00002.safetensors`
- `model.safetensors-00002-of-00002.safetensors`
- `model.safetensors.index.json`
- `tokenizer.json`, `tokenizer_config.json`, `vocab.json`, `merges.txt`
- `chat_template.jinja`
- `preprocessor_config.json`, `video_preprocessor_config.json`
- `LICENSE`, `NOTICE`

## Usage

Install a Transformers build that supports Qwen3.5, then load the model using
the standard Hugging Face APIs.

```python
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "ttrpg/mosslight-4b"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Briefly introduce yourself."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0], skip_special_tokens=True))
```

## Serving

Use serving frameworks only after confirming they support Qwen3.5 model classes
and the required multimodal processor files.

Example model identifier:

```bash
ttrpg/mosslight-4b
```

## Intended Use

Mosslight 4B is intended for experimentation with compact multimodal assistant
workflows, text generation, visual question answering, and local model serving.

## Limitations

- No independent benchmark results are published for this custom release yet.
- Behavior and safety characteristics should be evaluated for your target use
  case before deployment.
- This model inherits limitations from the Qwen3.5-4B base model and from the
  fine-tuning and merge process used for this release.

## Attribution

Mosslight 4B is a fine-tuned, merged derivative based on Qwen3.5-4B. Please
retain the Apache 2.0 license and attribution notices when redistributing this
model or derivatives of it.