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 <think>...</think> 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 <ser>...</ser> 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
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
API Integration: Dashboard

For Developers - Basic Usage

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

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

Continuous Pretraining: Extended training on reasoning-focused corpora
Advanced Reasoning Fine-tuning: Specialized training on intermediate thinking patterns
Multilingual Alignment: Cross-language reasoning consistency training
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:

<think>
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.
</think>

<ser>
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  
</ser>

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.

<think>
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.
</think>

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.

Hardware Type: H100 GPUs
days used: 16.2
Cloud Provider: Various
Compute Region: Multiple

Citation

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: <think>...</think> 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
Issues: Model Repository Issues

Benchmark

These benchmarks are done by @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

HelpingAI
/

Dhanishtha-2.0-preview-mlx