Dhanishtha-2.0-preview

Hey Product Hunt! 👋 We at HelpingAI were frustrated with reasoning models like DeepSeek-R1, Grok, and O3 — they take a lot of tokens, were slow and very expensive especially for startups and solopreneurs. While fine-tuning our old reasoning model, a bug accidentally made it generate think tag mid response. Instead of crashing, it got smarter. Intrigued, we went all-in and trained a model using this “mistake” — and it led to a breakthrough we call Intermediate Reasoning. We fine-tuned Qwen3-14B on 3 trillion tokens of multi-stage reasoning data. The result? A model that solves hard math problems, coding challenges, and logic riddles in 5x less time than DeepSeek, Grok, and O3. 👉 On one math problem: • Dhanishtha-2.0-preview: 45 seconds • DeepSeek: 280+ seconds 🤯 You can try it out now: 🌐 Chat & API – https://helpingai.co 🧠 Model (Open Source) – https://huggingface.co/HelpingAI... We’d love to hear your thoughts, feedback, or even crazy use cases you try it on! Cheers, Varun Gupta Co-founder, HelpingAI