qwen3-1.7b-nothink-gguf

Qwen3-1.7B-nothink

A modified (not fine-tuned) version of Qwen3-1.7B with Chain-of-Thought (CoT) completely disabled for faster responses, since enable_thinking=False (or /no_think) is not perfect and does not completely prevent reasoning in certain contexts.
This model supports llama.cpp and other compatible apps using llama.cpp as the backend.

Modification Details

The model has been modified (just two tokens) to prevent spontaneous Chain-of-Thought reasoning:

Changes:

• lm_head[151667] (token <think>): Set to a zero vector
• lm_head[151668] (token </think>): Set to the weights of original <think>

Thus the model never outputs <think>.
Manual modification is used since fine-tuning may negatively affect some model capabilities (especially for very small models).

Effect:

The <think> token generated by models are automatically converted to </think>, thus the model will generate non-reasoning contents after </think>.

An example output using the user prompt What is the .safetensors format, with T=0.5:

</think>

The `.safetensors` format is a new file format introduced by Hugging Face for storing large models' weights in a more efficient and secure way. It was designed to address some of the limitations of traditional formats like `TensorFlow SavedModel` or `PyTorch TorchScript`.

### Key Features of `.safetensors`:

1. **Efficiency**:
   - Stores model weights as a binary file, using a compact format that is more efficient than text-based formats.
   - Uses a binary format (not JSON or XML), which is faster to read and write.

2. **Security**:
   - It avoids the potential security issues of loading models directly from files, especially in environments where sensitive data might be exposed.
   - It provides better protection against malicious content by using a more secure storage mechanism.

3. **Compatibility**:
   - The format is compatible with Hugging Face's `transformers` library and other frameworks that support it.
   - It supports both CPU and GPU versions of models, as well as different model sizes and architectures.

4. **Ease of Use**:
   - Models can be saved in `.safetensors` format using the `save_pretrained()` method from the Hugging Face Transformers library.
   - The format is also supported by tools like `transformers` and `torch`.

...

The modification tool

See ekcbw/llm-nothink.