Shaik1903/ThinkLess-2B

ThinkLess-2B: Efficient and Accurate Reasoning

ThinkLess-2B, developed by Shaik1903, is a 2.3 billion parameter model based on Qwen3.5-2B, engineered to provide more accurate answers with significantly fewer reasoning tokens. It achieves this through a two-stage post-training process: Supervised Fine-Tuning (SFT) on the model's own shortest correct solutions (self-distillation) and a subsequent GRPO (Generative Reinforcement Learning with Policy Optimization) run with an accuracy-and-length reward.

Key Capabilities & Performance:

• Reduced Reasoning Tokens: Uses 57-82% fewer reasoning tokens on math and science benchmarks compared to the base Qwen3.5-2B.
• Improved Accuracy: Demonstrates higher accuracy on GSM8K (+3.7%), MATH-500 (+5.3%), and GPQA-Diamond (+8.6%) benchmarks.
• Faster Inference: Achieves 3.4x faster median request latency with vLLM and MTP speculative decoding due to shorter outputs.
• Robustness: Significantly reduces instances of answers being cut off mid-thought, ensuring complete responses.
• Quantization: Offers a ThinkLess-2B-FP8 version for half the memory footprint at near-identical accuracy.

Good for:

• Mathematical and Scientific Problem Solving: Excels in tasks requiring precise, step-by-step reasoning.
• Resource-Constrained Environments: Its efficiency in token usage and faster inference make it suitable for applications where speed and cost are critical.
• Applications requiring concise answers: Ideal for scenarios where verbose reasoning is undesirable, as it prunes redundant thought processes.

Limitations:

• For olympiad-level math problems, the ThinkLess-2B-SFT variant might offer higher accuracy, trading some reasoning length for peak performance.
• 4-bit AWQ quantization negatively impacts this model's performance and is not recommended; FP8 is the preferred compression method.