1. Introduction
We're introducing GRM-3.2-Cliff, our intermediate model built for long-horizon agentic tasks and extremely difficult reasoning problems in local environments. GRM-3.2-Cliff marks a substantial leap in long-horizon task capability over its predecessor, GRM-2.5-Plus, and is designed to serve as a dependable engine for complex, multi-step local workflows.
The model is purpose-built for long-horizon agentic tasks and problems that are simply hard — difficult coding challenges, advanced mathematics, and rigorous logical reasoning. GRM-3.2-Cliff aims to sustain coherent, goal-directed behavior over extended interactions while remaining optimized for resource-constrained hardware, making it ideal for developers and researchers who need local execution without sacrificing multi-step planning and self-correction performance.
2. Key Capabilities
- Long-Horizon Agentic Mastery: GRM-3.2-Cliff is specifically optimized to maintain coherence, planning quality, and task fidelity across long, multi-step agentic workflows, representing a major upgrade over GRM-2.5-Plus.
- Local Workflow Efficiency: Engineered to run smoothly in lower GPU environments while delivering high-tier reasoning performance.
- Elite Reasoning on Hard Problems: Strong performance on difficult coding, advanced mathematics, and logical reasoning tasks with careful, structured step-by-step problem-solving.
- Robust Coding Ability: Handles complex, multi-file coding tasks, debugging, refactoring, and long-running terminal sessions locally.
- Consistent Logical Reasoning: Built to reason carefully through multi-constraint logic problems without losing track of intermediate steps over extended execution runs.
3. Performance
GRM-3.2-Cliff is designed as our premier mid-sized model for local, long-horizon agentic work. It builds directly on the strengths of GRM-2.5-Plus while targeting common edge-case failures in smaller models — contextual drift, multi-step degradation, and loss of initial goal states — delivering strong reliability across extended sessions.
Detailed Benchmarks
| GRM-3.2-Cliff | GRM-2.5-Plus | GPT-5.6-Luna | Sonnet 5 | Gemini 3 Pro | |
|---|---|---|---|---|---|
| Knowledge & STEM | |||||
| Multidisciplinary knowledgeMMLU-Pro | 83.3 | 84.2 | — | — | 89.8 |
| Scientific reasoningGPQA Diamond | 82.4 | 82.7 | 92.3 | — | 91.9 |
| Reasoning & Coding | |||||
| Competitive codingLiveCodeBench v6 | 69.3 | 67.2 | — | — | 82.9 |
| General Agent | |||||
| Agentic codingSWE-bench Verified | 70.3 | — | — | 85.2 | 76.2 |
| Real-world software engineeringSWE-bench Pro | 43.4 | 35.6 | 62.7 | 63.2 | — |
| Agentic terminal codingTerminal-Bench 2.1 (Terminus-2) | 45.3 | 27.8 | 84.7 | 80.4 | — |
| Repo-level code generationNL2Repo | 28.5 | — | — | — | — |
Scores are taken from each provider's own published model card, blog post, or system card where available; "—" indicates a score was not publicly reported by that provider at the time of writing. Different labs may use different agent scaffolds when reporting SWE-bench and Terminal-Bench results, so cross-provider comparisons should be read with that caveat.
4. Family
The GRM-3.2 family is available in various sizes to suit every use case.
| Model | Domain |
|---|---|
| Sky35B-A3B | Flagship model for long-horizon tasks |
| Cliff9B | Capable model for low GPU environments |
| Turf1.2B | Lightweight model for practical reasoning |
5. Architecture
GRM-3.2-Cliff is built on the Ornith-1.0-9B architecture, a 9B-parameter model optimized for long-horizon agentic workflows, complex coding tasks, advanced mathematics, and logical reasoning, structured to run efficiently in low-to-mid GPU hardware environments.
GRM-3.2-Cliff is developed by OrionLLM and released under the Apache 2.0 License.