🧠 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.
