OrionLLM/GRM-3.2-Cliff
OrionLLM/GRM-3.2-Cliff is a 9 billion parameter model built on the Ornith-1.0-9B architecture, specifically optimized for long-horizon agentic tasks and complex reasoning problems. It excels at difficult coding challenges, advanced mathematics, and rigorous logical reasoning, maintaining coherence over extended interactions. Designed for local execution, it runs efficiently in lower GPU environments while delivering high-tier performance for multi-step planning and self-correction. This model is ideal for developers and researchers requiring robust local agentic capabilities.
Loading preview...
OrionLLM/GRM-3.2-Cliff Overview
GRM-3.2-Cliff is a 9 billion parameter model developed by OrionLLM, designed for long-horizon agentic tasks and extremely difficult reasoning problems in local environments. Built on the Ornith-1.0-9B architecture, it represents a significant advancement over its predecessor, GRM-2.5-Plus, in sustaining coherent, goal-directed behavior over extended interactions.
Key Capabilities
- Long-Horizon Agentic Mastery: Optimized for maintaining coherence, planning quality, and task fidelity across complex, multi-step agentic workflows.
- Local Workflow Efficiency: Engineered to operate smoothly in lower GPU environments without sacrificing high-tier reasoning performance.
- Elite Reasoning: Demonstrates strong performance on challenging coding, advanced mathematics, and logical reasoning tasks through careful, structured step-by-step problem-solving.
- Robust Coding Ability: Capable of handling complex, multi-file coding tasks, debugging, refactoring, and long-running terminal sessions locally.
Performance Highlights
GRM-3.2-Cliff is positioned as a premier mid-sized model for local agentic work, addressing common failures in smaller models such as contextual drift and multi-step degradation. It achieves competitive scores on various benchmarks, including 70.3 on SWE-bench Verified and 43.4 on SWE-bench Pro, indicating strong reliability across extended sessions for complex tasks. The model also shows a notable 28.5 on NL2Repo, a benchmark where other leading models did not report scores.
Good For
This model is particularly well-suited for developers and researchers who need a dependable engine for complex, multi-step local workflows, especially those involving difficult coding, advanced mathematics, and rigorous logical reasoning, while operating within resource-constrained hardware environments.