We’re the team behind HelpingAI, and we’ve been experimenting with ways to make LLM reasoning faster and more efficient, especially for complex tasks like math problems, coding questions, and logic puzzles.
During a fine-tuning session of our older reasoning model, a bug accidentally introduced a think tag mid-response. Rather than break the model, it created a chain-of-thought style intermediate step — and the model got better. That bug inspired us to explore what we now call Intermediate Reasoning.
We scaled this up by fine-tuning Qwen3-14B on 3 trillion tokens of reasoning-specific data — curated for multistage reasoning tasks — and the results blew us away:
• 5x faster reasoning vs. models like DeepSeek-R1, Grok, and OpenChat.
• Significantly lower token usage and latency, ideal for startups and real-time use cases.
• Solves benchmark-level math and logic problems in seconds instead of minutes.
On a hard math benchmark, for example:
• Dhanishtha-2.0 solved it in 45s
• DeepSeek-R1 took 280+ seconds
You can try it here:
• Chat/API access: https://helpingai.co
• Model weights (open source): https://huggingface.co/HelpingAI/Dhanishtha-2.0-preview
We’d love feedback, ideas, and especially edge cases where you think it might break (or surprise you). Happy to answer any questions and go into training data, fine-tuning strategy, evals, or infra!
Thanks,
Varun Gupta
Co-founder, HelpingAI