Dhanishtha-2.1: Introducing Adaptive Effort Intelligence

The problem with one-size-fits-all thinking

Traditional language models typically operate at a single cognitive intensity. Whether answering "What's 2+2?" or solving complex mathematical proofs, they apply the same computational effort. It's like using a sledgehammer for every task—regardless of whether you're cracking a nut or breaking concrete.

Dhanishtha-2.1 changes this paradigm.

Four levels of effort

At the heart of Dhanishtha-2.1 is our Adaptive Effort Architecture. The model supports four discrete effort modes so it can match compute to task complexity:

• No effort — Instant responses with essentially no computational overhead. Unlike some approaches, we avoid sending empty "think" blocks, which eliminates token waste.
• Low effort — Fast, efficient replies for straightforward queries (this was the default behavior of the original Dhanishtha).
• Medium effort — Balanced, moderate-depth reasoning for typical, slightly complex tasks.
• High effort — Deep, extended reasoning for problems that require comprehensive analysis.

This isn't only about speed. It's intelligent resource allocation: the model learns when to think harder and when to respond instinctively.

Revolutionary training approach

Dataset generation — teaching models to think at different depths

We pioneered a dataset methodology that teaches the model how to respond at each effort level:

• Every training example contains four distinct answers — one per effort level.
• Smaller Dhanishtha-2.1 models were used to generate datasets for larger models, creating a self-improving pipeline.
• Our custom agentic framework used ensembles of Dhanishtha-2.0 models (5×) working in concert to generate high-quality training data.

Training innovations

Key technical choices we made to push performance:

• Most models were trained for a single epoch (the 14B variant trained for 3 epochs).
• Extremely small learning rate: 1e-37, for very fine-grained updates.
• Custom reinforcement-learning method: a modified GRPO (Gradient Reward Policy Optimization) adapted specifically for the 14B model.
• The 0.6B model reused and improved the regenerated dataset from Dhanishtha 2.0, extended to cover all four effort levels.

What this means for developers

Dhanishtha-2.1 is more than an incremental improvement — it rethinks how models allocate compute. Practical benefits for developers include:

• Build applications that automatically scale thinking effort based on query complexity.
• Reduce inference costs by up to ~80% on simple queries.
• Maintain peak performance for complex reasoning tasks by escalating effort only when needed.
• Deploy adaptive AI that behaves more like human problem-solving: efficient on the simple stuff, thorough on the hard stuff.

The road ahead

Dhanishtha-2.1 is a first step toward our larger release, "Dhanishtha Max," scheduled for 26 March 2025. The Max model will include five effort modes and further improvements to adaptivity and efficiency.

---

Published by the Dhanishtha research team.