1. Introduction
We're introducing GRM-3.2-Turf, our lightweight model built for difficult reasoning problems and general conversation in local environments. GRM-3.2-Turf marks a substantial leap over its predecessor, GRM-2.6-Air-Opus, and is designed to serve as a dependable engine for low-resource devices with low processing power.
The model is purpose-built to deliver efficient execution without sacrificing structured reasoning capability, making it ideal for mobile devices, embedded systems, and local deployments where memory and compute constraints are critical.
2. Key Capabilities
- On-Device Efficiency: Engineered to run smoothly on low-resource hardware with minimal memory footprint and fast inference latency.
- Enhanced Local Reasoning: Substantial leap over GRM-2.6-Air-Opus in structured reasoning, problem-solving, and general conversation tasks.
- High-Fidelity Instruction Following: Exceptional capability in handling constrained prompts, complex system instructions, and precise response formatting.
- Robust Tool Use: Strong performance in tool calling and function execution, enabling agentic workflows in lightweight environments.
3. Performance
GRM-3.2-Turf is designed as our premier lightweight model for local execution. It builds directly on the strengths of GRM-2.6-Air-Opus while setting new benchmarks for sub-2B reasoning and instruction-following capability on edge hardware.
Detailed Benchmarks
| GRM-3.2-Turf | LFM2.5-1.2B-Thinking | Qwen3.5-2B | Gemma-4-E2B | |
|---|---|---|---|---|
| Knowledge & STEM | ||||
| Multidisciplinary knowledgeMMLU-Pro | 56.2 | 49.65 | 66.5 | 60.0 |
| Scientific reasoningGPQA Diamond | 42.4 | 37.86 | 51.6 | 43.4 |
| Instruction Following & Function Calling | ||||
| Instruction followingIFEval | 91.2 | 88.42 | 78.6 | — |
| Complex instruction followingIFBench | 46.8 | 44.85 | 41.3 | — |
| Function callingBFCL v3 | 59.3 | 56.97 | — | — |
Scores are taken from each provider's own published model card, blog post, or evaluation benchmark suite.
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-Turf is built on the LiquidAI/LFM2.5-1.2B-Thinking base architecture, a 1.2B-parameter model optimized for reasoning, general conversation, and tool use, specifically tailored for efficient deployment on resource-constrained hardware.
GRM-3.2-Turf is developed by OrionLLM and released under the Apache 2.0 License.