---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3-4B
pipeline_tag: text-generation
library_name: transformers
tags:
- moe
- moderately abliterated variant
- text-generation-inference
---

![FMjPew6Vjrp4FvKe1Uz_T.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/cMq6BmD-1cEiOngYBL3Wh.png)

# **Qwen3-4B-ft-bf16**

> **Qwen3-4B-ft-bf16** is a fine-tuned, moderately abliterated version of the Qwen3-4B model. Designed for **enhanced context awareness** and **controlled expressiveness**, this model balances precision with creativity across a wide range of tasks—from complex reasoning to natural dialogue, code generation, and multilingual understanding.

### Key Features:

- **Improved Context Awareness**  
  Retains and utilizes long-range contextual information effectively, making it ideal for long-form conversations, document understanding, and summarization tasks.

- **Moderate Abliteration**  
  Introduces measured behavioral flexibility that enhances creativity and adaptability while maintaining reliability, alignment, and safety in outputs.

- **Dual Thinking Modes**  
  Supports dynamic switching between *thinking* mode (for math, logic, and coding) and *non-thinking* mode (for general-purpose conversations), ensuring optimal task matching.

- **Multilingual Mastery**  
  Excels in over 100 languages and dialects for translation, multilingual chat, and cross-lingual reasoning.

- **Tool-Ready Agent Capabilities**  
  Designed to integrate with tool APIs and complex workflows, with consistent performance in both thinking and non-thinking contexts.

---

## Quickstart with Hugging Face Transformers🤗

```bash
pip install transformers==4.51.3
pip install huggingface_hub[hf_xet]
```

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Qwen3-4B-ft-bf16"

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Define input
prompt = "Describe how renewable energy impacts economic development."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate output
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# Parse thinking content
try:
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip()
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip()

print("thinking content:", thinking_content)
print("content:", content)
```

---

## Best Practices

- **Sampling Settings**:
  - *Thinking mode*: `temperature=0.6`, `top_p=0.95`, `top_k=20`
  - *Non-thinking mode*: `temperature=0.7`, `top_p=0.8`, `top_k=20`

- **Token Length**:
  - Standard: `32768 tokens`
  - Extended Reasoning Tasks: `up to 38912 tokens`

- **Prompt Design**:
  - **Math Problems**: Add `"Please reason step by step, and put your final answer within \boxed{}."`
  - **MCQs**: Format answers as `{"answer": "B"}` for easy parsing.
  - **Multi-turn**: Omit thinking logs in conversation history for cleaner context.