OrionLLM/LRM-3.2
OrionLLM/LRM-3.2 is a 27 billion parameter reasoning model fine-tuned from Qwen3.6-27B with a 32768 token context length. Developed by Orion Research, it specializes in highly compressed, direct chain-of-thought reasoning, significantly reducing token usage during internal thought processes. This model maintains full depth of reasoning and high-quality final answers while dramatically lowering inference latency and token spend, making it ideal for efficient agentic and multi-step tasks.
Loading preview...
OrionLLM/LRM-3.2: Efficient Reasoning with Compressed Thought
LRM-3.2 is a 27 billion parameter model developed by Orion Research, fine-tuned from Qwen3.6-27B. Its core innovation lies in its ability to perform deep, structured chain-of-thought reasoning using significantly fewer tokens than traditional models. This is achieved by stripping out verbose narrative padding from the internal thinking process, while preserving the full depth of deliberation and ensuring final answers remain high-quality and natural.
Key Capabilities
- Adaptive Reasoning Depth: The model dynamically adjusts the length of its internal thought process based on task complexity, providing concise reasoning for simple problems and detailed deliberation for complex ones.
- Unaltered Output Quality: The compression is applied exclusively to the reasoning trace; the final generated answers are equivalent in quality and naturalness to those from models with verbose thinking.
- Optimized Inference: By drastically shortening the internal 'think' blocks, LRM-3.2 achieves lower latency and reduced token consumption per response without sacrificing capability.
- Sustained Performance: It maintains dense, on-task reasoning even during extended agentic workflows and multi-step tasks, avoiding repetitive verbal habits.
Performance and Efficiency
LRM-3.2 demonstrates strong performance in reasoning and coding benchmarks, scoring 95.8 on GSM8K and 86.9 on HumanEval. A key differentiator is its efficiency: for a complex coding task, LRM-3.2 produced a complete reasoning trace in just 33 tokens, compared to 6,539 tokens for its base model, Qwen3.6-27B, while yielding the same correct solution. This represents a 198x reduction in thinking token usage.
Training Details
LRM-3.2 was fine-tuned using the grug-think and grug-think-v3-10k datasets. The training applied a think-only loss on trajectory data, specifically teaching the model to adopt a compressed reasoning style without impacting the quality of its final answers. This approach is inspired by ProCreations/grug-27b, which pioneered padding-free thinking.