OrionLLM/GRM-3.2-Turf
GRM-3.2-Turf is a 1.2 billion parameter language model developed by OrionLLM, built upon the LiquidAI/LFM2.5-1.2B-Thinking base architecture. This model is specifically engineered for difficult reasoning problems and general conversation in local, low-resource environments. It excels in on-device efficiency, enhanced local reasoning, high-fidelity instruction following, and robust tool use, making it ideal for mobile devices and embedded systems.
Loading preview...
Overview
OrionLLM's GRM-3.2-Turf is a 1.2 billion parameter model designed for efficient execution on resource-constrained hardware. Built on the LiquidAI/LFM2.5-1.2B-Thinking architecture, it focuses on solving difficult reasoning problems and facilitating general conversation in local environments. This model represents a significant improvement over its predecessor, GRM-2.6-Air-Opus, by offering enhanced structured reasoning capabilities without sacrificing efficiency.
Key Capabilities
- On-Device Efficiency: Optimized for smooth operation on low-resource hardware with minimal memory footprint and fast inference.
- Enhanced Local Reasoning: Demonstrates a substantial leap in structured reasoning, problem-solving, and general conversation tasks compared to previous versions.
- High-Fidelity Instruction Following: Capable of handling complex system instructions, constrained prompts, and precise response formatting.
- Robust Tool Use: Exhibits strong performance in tool calling and function execution, supporting agentic workflows in lightweight settings.
Performance Highlights
GRM-3.2-Turf sets new benchmarks for sub-2B models in reasoning and instruction-following on edge hardware. It outperforms LFM2.5-1.2B-Thinking across several key metrics:
- MMLU-Pro: Achieves 56.2 (vs. 49.65)
- GPQA Diamond: Scores 42.4 (vs. 37.86)
- IFEval: Reaches 91.2 (vs. 88.42)
- BFCL v3: Scores 59.3 (vs. 56.97)
Ideal Use Cases
GRM-3.2-Turf is particularly well-suited for applications requiring powerful language capabilities on devices with limited processing power and memory, such as mobile devices, embedded systems, and other local deployments.