# FastApply-1.5B-v1.0 **🚀 Update September 2025:** For production-grade throughput, we use *[Morph](https://morphllm.com)* (the hosted Fast Apply API powering [SoftGen AI](https://softgen.ai/)). - Morph hits **~10,500 tok/s** even on huge token diffs **44k input** - Larger model trained on millions of examples and tuned for accuracy. > Stable inference, large free tier, highly recommended if you need serious speed in prod. [Github: kortix-ai/fast-apply](https://github.com/kortix-ai/fast-apply) [Dataset: Kortix/FastApply-dataset-v1.0](https://huggingface.co/datasets/Kortix/FastApply-dataset-v1.0) [Try it now on 👉 Google Colab](https://colab.research.google.com/drive/1BNCab4oK-xBqwFQD4kCcjKc7BPKivkm1?usp=sharing) ## Model Details ### Basic Information - **Developed by:** Kortix - **License:** apache-2.0 - **Finetuned from model:** [unsloth/Qwen2.5-Coder-1.5B-Instruct-bnb-4bit](https://huggingface.co/unsloth/Qwen2.5-Coder-1.5B-Instruct-bnb-4bit) ### Model Description FastApply-1.5B-v1.0 is a 1.5B model designed for instant code application, producing full file edits to power [SoftGen AI](https://softgen.ai/). It is part of the Fast Apply pipeline for data generation and fine-tuning Qwen2.5 Coder models. The model achieves high throughput when deployed on fast providers like Fireworks while maintaining high edit accuracy, with a speed of approximately 340 tokens/second. ## Intended Use FastApply-1.5B-v1.0 is intended for use in AI-powered code editors and tools that require fast, accurate code modifications. It is particularly well-suited for: - Instant code application tasks - Full file edits - Integration with AI-powered code editors like Aider and PearAI - Local tools to reduce the cost of frontier model output ## Inference template FastApply-1.5B-v1.0 is based on the Qwen2.5 Coder architecture and is fine-tuned for code editing tasks. It uses a specific prompt structure for inference: ``` <|im_start|>system You are a coding assistant that helps merge code updates, ensuring every modification is fully integrated.<|im_end|> <|im_start|>user Merge all changes from the snippet into the

 below.
- Preserve the code's structure, order, comments, and indentation exactly.
- Output only the updated code, enclosed within  and  tags.
- Do not include any additional text, explanations, placeholders, ellipses, or code fences.

{original_code}

{update_snippet}

Provide the complete updated code.<|im_end|>
<|im_start|>assistant
```

The model's output is structured as:

```
[Full-complete updated file]
```

## Additional Information

For more details on the Fast Apply pipeline, data generation process, and deployment instructions, please refer to the [GitHub repository](https://github.com/kortix-ai/fast-apply).

## How to Use

To use the model, you can load it using the Hugging Face Transformers library:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Kortix/FastApply-1.5B-v1.0", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Kortix/FastApply-1.5B-v1.0")

# Prepare your input following the prompt structure mentioned above
input_text = """<|im_start|>system
You are a coding assistant that helps merge code updates, ensuring every modification is fully integrated.<|im_end|>
<|im_start|>user
Merge all changes from the  snippet into the  below.
- Preserve the code's structure, order, comments, and indentation exactly.
- Output only the updated code, enclosed within  and  tags.
- Do not include any additional text, explanations, placeholders, ellipses, or code fences.

{original_code}

{update_snippet}

Provide the complete updated code.<|im_end|>
<|im_start|>assistant
"""

input_text = input_text.format(
    original_code=original_code,
    update_snippet=update_snippet,
).strip() 

# Generate the response
input_ids = tokenizer.encode(input_text, return_tensors="pt")
output = model.generate(input_ids, max_length=8192,)

response = tokenizer.decode(output[0][len(input_ids[0]):])
print(response)

# Extract the updated code from the response
updated_code = response.split("")[1].split("")[0]
```

## Evaluation:

![image/png](https://cdn-uploads.huggingface.co/production/uploads/650d7ecb23e8028a8970a203/_E6WVzuVABKB58QMx6c1c.png)