Hugging Face's logo

BrainWave-ML
/
ThoughtSwitch-V1-1.7b-Instruct

Safetensors
qwen3
Model card Files
xet
Community
ThoughtSwitch-V1-1.7b-Instruct / README.md
Shaleen123's picture
Shaleen123
Update README.md
3226a14 verified
preview code
|
raw
history blame contribute delete
4.88 kB

🧠 ThoughtSwitch V1 1.7B Instruct — A Mode-Adaptive Reasoning Language Model

Model ID: BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct
Architecture: Decoder-only transformer (GPT-style)
Parameters: 1.7 Billion
Capabilities: Dynamic "Thinking" vs. "Non-Thinking" mode-switching
Fine-Tuned for: Instruction-following

🚀 Overview

ThoughtSwitch V1 is a next-generation instruction-tuned language model that brings a new paradigm to text generation: Autonomous Cognitive Mode Switching.

It is capable of interpreting user prompts and switching between two distinct modes of behavior:

  • 🧠 Thinking Mode: Deep reasoning, logical step-by-step solutions, slow but deliberate outputs.
  • 💬 Non-Thinking Mode: Quick completions, casual replies, storytelling, and chat-like fluency.

Whether you're building reasoning agents, fast assistants, or multi-modal chains-of-thought applications, ThoughtSwitch adapts intelligently—so you don’t have to force the prompt.

🧠 Key Features

  • ✅ Autonomous Mode Switching
    Understands when to think deeply and when to generate fluently, based on prompt phrasing.

  • ✅ Instruction Tuned
    Trained to follow human-like instructions and align closely with user intent.

  • ✅ 1.7B Parameters
    Small enough for efficient inference, yet powerful for sophisticated reasoning.

  • ✅ Open Weights
    Fully accessible under a permissive license (specify in HF model card).

✨ Example Prompts

Prompt (Thinking Mode): "Think step by step to solve this math problem: What is 17 multiplied by 23?"

→ Reasoned output with intermediate steps and justification.

Prompt (Non-Thinking Mode): "Write a quick sci-fi story about a robot discovering love."

→ Smooth, creative storytelling without unnecessary reasoning.

🔧 Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct") model = AutoModelForCausalLM.from_pretrained("BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct")

prompt = "Think step by step: Why does ice float on water?" inputs = tokenizer(prompt, return_tensors="pt") outputs = model.generate(**inputs, max_new_tokens=150) print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🧪 Intended Use Cases

  • 🧠 Reasoning Agents — For multi-hop question answering, logical puzzles, or decision support.
  • 📚 Tutoring & Education — Adaptive explanations that vary depth based on student prompts.
  • 🤖 Conversational AI — More natural and flexible interactions with variable "thinking effort".
  • ✍️ Creative Writing — Generate stories, poems, and ideas with or without deep context.

⚠️ Limitations

  • Like all LLMs, it may hallucinate or generate biased content.
  • Mode switching is probabilistic, not guaranteed—prompt clearly for best results.
  • Performance may vary outside of English or unfamiliar domains.

📈 Performance (Unofficial Benchmarks)

Task Performance
Commonsense Reasoning ✅ Strong
Instruction Following ✅ Strong
Fast Casual Generation ✅ Very Strong
Math (Step-by-Step) ⚠️ Moderate
Factual QA ⚠️ May hallucinate

🛠️ Model Details

  • Architecture: GPT-style decoder (causal LM)
  • Training: Custom pretraining with hybrid reasoning/non-reasoning dataset
  • Instruction Fine-Tuning: Yes, using curated prompt-response pairs
  • Token Limit: 2048 tokens (extendable with rope scaling)

🔍 Quantized Version

Looking for fast inference?

Check out the GGUF-quantized version (by @mradermacher) for compatibility with llama.cpp, KoboldAI, and other lightweight runtimes.

📄 Citation

If you use this model in your research or application, please cite it as:

@misc{thoughswitch2025, title={ThoughtSwitch V1 1.7B Instruct: A Mode-Adaptive Reasoning Language Model}, author={BrainWave-ML}, year={2025}, howpublished={\url{https://huggingface.co/BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct}} }

💬 Contact

For issues, feedback, or collaboration:

🙌 Acknowledgments

Developed by the team at BrainWave-ML. Inspired by the question:
“What if language models could choose when to think?”

ThoughtSwitch: Think when you need to. Generate when you don't.