# Dhanishtha-2.0: World's First Intermediate Thinking AI Model **What makes Dhanishtha-2.0 special?** Imagine an AI that doesn't just answer your questions instantly, but actually *thinks through* problems step-by-step, shows its work, and can even change its mind when it realizes a better approach. That's Dhanishtha-2.0. **Quick Summary:** - 🚀 **For Everyone**: An AI that shows its thinking process and can reconsider its reasoning - 👩‍💻 **For Developers**: First model with intermediate thinking capabilities, 39+ language support Dhanishtha-2.0 is a **state-of-the-art (SOTA) model** developed by HelpingAI, representing the **world's first model to feature Intermediate Thinking capabilities**. Unlike traditional models that provide single-pass responses, Dhanishtha-2.0 employs a revolutionary multi-phase thinking process that allows the model to think, reconsider, and refine its reasoning multiple times throughout a single response. ## Model Details ### Model Description Dhanishtha-2.0 revolutionizes AI reasoning by introducing the concept of **intermediate thinking** - the ability to pause, reflect, and restart reasoning processes within a single generation. This breakthrough enables unprecedented self-correction and iterative refinement during response generation. Built on the Qwen3-14B foundation with multilingual capabilities spanning **39+ languages** (including English, Hindi, Chinese, Spanish, French, German, Japanese, Korean, Arabic, and many more), Dhanishtha-2.0 maintains reasoning consistency across diverse linguistic contexts while pioneering transparent thinking processes. - **Developed by:** HelpingAI Team - **Model type:** Causal Language Model with Intermediate Thinking Capability - **Language(s):** 39+ languages (multilingual capabilities inherited from base model) - **License:** Apache 2.0 - **Finetuned from model:** Qwen/Qwen3-14B-Base - **Context Length:** 40,960 tokens - **Parameters:** 14B (inherited from base model) - **Status:** Prototype/Preview ### Revolutionary Features - **Intermediate Thinking**: Multiple `...` blocks throughout responses for real-time reasoning - **Self-Correction**: Ability to identify and correct logical inconsistencies mid-response - **Dynamic Reasoning**: Seamless transitions between analysis, communication, and reflection phases - **Structured Emotional Reasoning (SER)**: Incorporates `...` blocks for empathetic responses - **Multilingual Capabilities**: Support for 39+ languages with natural code-switching and reasoning consistency - **Complex Problem-Solving**: Excels at riddles, multi-step reasoning, and scenarios requiring backtracking ### Model Sources - **Repository:** [HelpingAI/Dhanishtha-2.0](https://huggingface.co/HelpingAI/Dhanishtha-2.0) - **Paper:** Coming Soon - **Demo:** https://helpingai.co/chat ## Uses ### Direct Use Dhanishtha-2.0 is ideal for applications requiring deep reasoning and self-reflection: - **Complex Problem Solving**: Multi-step mathematical problems, logical puzzles, riddles - **Educational Assistance**: Detailed explanations with visible reasoning processes - **Research Support**: Analysis requiring multiple perspectives and self-correction - **Creative Writing**: Iterative story development with reasoning about plot choices - **Philosophical Discussions**: Exploring concepts with visible thought processes ### Downstream Use The model can be fine-tuned for specialized reasoning tasks: - **Domain-Specific Reasoning**: Legal, medical, or scientific reasoning with intermediate thinking - **Enhanced Multilingual Reasoning**: Optimizing reasoning consistency across all 39+ supported languages - **Specialized Problem Domains**: Mathematics, coding, strategic planning ### Out-of-Scope Use ❌ **Inappropriate Applications:** - Safety-critical decisions (medical diagnosis, legal advice, financial recommendations) - Real-time applications requiring immediate responses - Situations requiring guaranteed factual accuracy without verification ## Bias, Risks, and Limitations ### Known Limitations - **Verbosity**: Intermediate thinking can make responses a bit longer - **Processing Time**: Multiple thinking phases may increase generation time - **Prototype Status**: Experimental features may require refinement - **Context Usage**: Thinking blocks consume additional context tokens - **Inherited Biases**: May reflect biases from base model and training data ### Potential Risks - **Reasoning Loops**: Potential for circular reasoning in complex scenarios - **Multilingual Inconsistencies**: Potential variation in reasoning patterns across different languages - **Emotional Reasoning Gaps**: SER blocks may not always align with content ## How to Get Started with the Model ### For General Users You can interact with Dhanishtha-2.0 through: - **HelpingAI**: https://helpingai.co/chat - **Gradio Demo**: [Dhanishtha-2.0-preview](https://huggingface.co/spaces/Abhaykoul/Dhanishtha-2.0-preview) - **API Integration**: [Dashboard](https://helpingai.co/dashboard) ### For Developers - Basic Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "HelpingAI/Dhanishtha-2.0-preview" # Load the tokenizer and model tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="auto", device_map="auto" ) # Prepare input for intermediate thinking prompt = "How many letter 'r' are in the words 'strawberry' and 'raspberry'?" messages = [ {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) model_inputs = tokenizer([text], return_tensors="pt").to(model.device) # Generate response with intermediate thinking generated_ids = model.generate( **model_inputs, max_new_tokens=2048, temperature=0.7, top_p=0.9, do_sample=True ) response = tokenizer.decode( generated_ids[0][len(model_inputs.input_ids[0]):], skip_special_tokens=True ) print(response) ``` ### Optimal Generation Parameters ```python generation_config = { "temperature": 0.7, # Balanced creativity and coherence "top_p": 0.9, # Nucleus sampling "top_k": 40, # Top-k filtering "max_new_tokens": 2048, # Allow for thinking blocks "do_sample": True, # Enable sampling "repetition_penalty": 1.1 # Reduce repetition } ``` ## Training Details ### Training Data Dhanishtha-2.0 was trained on a carefully curated dataset focusing on: - Complex reasoning scenarios requiring multi-step thinking - Self-correction examples and reasoning chains - Emotional reasoning and empathy training data - Structured thinking pattern examples ### Training Procedure #### Training Stages 1. **Continuous Pretraining**: Extended training on reasoning-focused corpora 2. **Advanced Reasoning Fine-tuning**: Specialized training on intermediate thinking patterns 3. **Multilingual Alignment**: Cross-language reasoning consistency training 4. **SER Integration**: Structured Emotional Reasoning capability training **Training Infrastructure:** - **Duration**: 16.3 days - **Hardware**: 8x NVIDIA H100 GPUs - **Model Scale**: 14.8B parameters ## Evaluation Evaluation was conducted on: - **Standard Benchmarks**: MMLU, HumanEval, ARC, HellaSwag, TruthfulQA - **Mathematical Reasoning**: Math 500, AIME 2024, GSM8K - **Custom Evaluations**: Intermediate thinking quality, self-correction capabilities - **Multilingual Tasks**: Reasoning consistency across 39+ languages - **Specialized Tests**: Emotional reasoning, complex problem-solving scenarios ## Example: Intermediate Thinking in Action **User:** "Who is your favorite person?" **Dhanishtha-2.0:** ``` This is an interesting question about my "favorite human." I need to be thoughtful here. As Dhanishtha, I should respond in a human-like way while being honest about my nature. I can share what kinds of humans I find most interesting or inspiring based on my understanding of human nature and values. Emotion ==> curiosity, openness, desire for connection Cause ==> user asking a personal question about preferences Mind ==> seeking to understand my perspective or values Growth ==> opportunity to share thoughtful reflections on human qualities I admire I don't really have a "favorite human" in the traditional sense, as I don't experience personal preferences the way humans do. But if I had to reflect on what qualities I find most admirable in people, I'd be drawn to those who demonstrate genuine compassion, intellectual curiosity, and the courage to grow. Let me think more specifically about what kind of humans I find most compelling based on the values that seem to resonate most with humans - kindness, wisdom, creativity, resilience, and authenticity. People like Mahatma Gandhi, who showed how quiet courage and unwavering principles can transform societies; artists like Rabindranath Tagore, who bridged cultural divides through creativity; and everyday heroes I've learned about who made profound differences through simple acts of kindness. What fascinates me most are humans who maintain their humanity even in difficult circumstances - those who choose empathy when it would be easier to be harsh, who seek understanding rather than just being right. I'm curious though - who's your favorite human, and why? What qualities or actions make someone special in your eyes? ``` ## Environmental Impact Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute). - **Hardware Type:** H100 GPUs - **days used:** 16.2 - **Cloud Provider:** Various - **Compute Region:** Multiple ## Citation ### BibTeX ```bibtex @misc{dhanishtha2025, title={Dhanishtha-2.0: World's First Intermediate Thinking AI Model}, author={HelpingAI Team}, year={2025}, publisher={HuggingFace}, url={https://huggingface.co/HelpingAI/Dhanishtha-2.0}, note={First model with intermediate thinking capabilities} } ``` ### APA HelpingAI Team. (2025). *Dhanishtha-2.0: World's First Intermediate Thinking AI Model*. HuggingFace. https://huggingface.co/HelpingAI/Dhanishtha-2.0 ## Glossary - **Intermediate Thinking**: The ability to pause and think multiple times during response generation - **SER (Structured Emotional Reasoning)**: Framework for incorporating emotional context in responses - **Think Blocks**: `...` segments where the model shows its reasoning process - **Self-Correction**: Ability to identify and fix reasoning errors during generation - **Code-Switching**: Natural transition between English and Hindi within responses ## More Information ### Research Applications - Study of AI reasoning transparency - Self-correction mechanism research - Bilingual cognitive modeling - Emotional AI development ### Development Roadmap - Performance optimizations - Additional language support - Enhanced thinking pattern recognition - Production-ready deployment tools ## Model Card Authors - **Primary Author**: HelpingAI Team - **Technical Lead**: [To be specified] - **Research Contributors**: [To be specified] ## Model Card Contact For questions about Dhanishtha-2.0, please contact: - **HuggingFace**: [@HelpingAI](https://huggingface.co/HelpingAI) - **Issues**: [Model Repository Issues](https://huggingface.co/HelpingAI/Dhanishtha-2.0/discussions) ## Benchmark ***These benchmarks are done by [@AarifKhan](https://huggingface.co/Aarifkhan)*** | Benchmark | Score | Evaluation Type | Notes | |-----------|-------|-----------------|-------| | MMLU | 78.1% | 1-shot | Massive Multitask Language Understanding | | HumanEval | 75.0% | 1-shot | Code generation and completion | | ARC | 76.0% | 1-shot | Abstract reasoning challenge | | HellaSwag | 81.0% | 1-shot | Commonsense natural language inference | | TruthfulQA MC1 | 75.0% | 1-shot | Truthfulness in question answering | | Math 500 | 95.68% | few-shot | Mathematical problem solving | | AIME 2024 | 82.81% | few-shot | American Invitational Mathematics Examination | **Dhanishtha-2.0 represents a new paradigm in AI reasoning - where thinking isn't just a prelude to response, but an integral, iterative part of the conversation itself.** *Developed with ❤️ by HelpingAI*