GRM-2.5
Collection
Reasoning models for complex reasoning, challenging tasks, and all kinds of chat and everyday use. • 2 items • Updated • 2
GRM-2.5-Air
1. Introduction
GRM-2.5-Air is a 0.8B-parameter reasoning model built for general-purpose local AI. It is designed to deliver strong performance across a wide range of tasks while remaining efficient and accessible for very small devices.
The model is optimized for structured reasoning, helping it produce more accurate, coherent, and reliable responses on complex problems. GRM-2.5-Air aims to combine strong reasoning ability, practical usability, and efficient deployment in a compact form factor.
2. Key Capabilities
- Strong Reasoning for Everyday and Advanced Tasks: GRM-2.5-Air is built to handle both daily conversations and more demanding reasoning workloads with clarity and consistency.
- Optimized for Local Deployment: GRM-2.5-Air is designed for accessible inference across a broad range of hardware, making it a practical choice for users who want capable AI running locally.
3. Performance
GRM-2.5-Air is designed to be a highly capable option for local AI use across many scenarios. It performs well in complex reasoning tasks, everyday chat, coding, and agentic workflows, while maintaining the efficiency expected from a compact 0.8B model.
Its focus is not only raw capability, but also practical intelligence: strong reasoning, stable long-context behavior, and usability on consumer hardware.
| GRM-2.5-Plus (Closed) | GRM-2.5 | GRM-2.5-Air | GRM-7B | GRM-1.5B | |
|---|---|---|---|---|---|
| Knowledge & STEM | |||||
| MMLU-Pro | 84.2 | 80.1 | 43.6 | -- | -- |
| GPQA Diamond | 82.7 | 76.7 | 12.5 | 53.7 | 29.5 |
| Instruction Following | |||||
| IFEval | 91.8 | 90.2 | 44.5 | -- | -- |
| MultiChallenge | 56.5 | 49.8 | 19.3 | -- | -- |
| Reasoning & Coding | |||||
| HMMT Feb 25 | 84.4 | 75.2 | -- | 42.7 | 27.3 |
| HMMT Nov 25 | 83.2 | 77.2 | -- | -- | -- |
| LiveCodeBench v6 | 67.2 | 56.9 | -- | 51.7 | 39.4 |
| Agent | |||||
| TAU2-Bench | 80.5 | 80.2 | 11.6 | -- | -- |
| DeepPlanning | 18.6 | 17.9 | -- | -- | -- |
| OSWorld-Verified | 42.4 | 36.0 | -- | -- | -- |
4. Family
The GRM-2.5 family is available in various sizes to suit every case.
| Model | Size | Domain |
|---|---|---|
| GRM-2.5-Plus | 9B | Closed model for research and agent purposes |
| GRM-2.5 | 4B | Powerful on-device deployment for difficult tasks |
| GRM-2.5-Air | 0.8B | Any-device deployment for everyday chat |
5. Architecture
GRM-2.5 is built on the Qwen3.5 architecture and is optimized for complex tasks, agent environments, and everyday chat.
GRM-2.5 applies the same principle to a stronger, larger foundation, resulting in a model that punches above its weight class on structured reasoning tasks while remaining deployable on consumer hardware.
GRM-2.5 is developed by OrionLLM and released under the Apache 2.0 License.
- Downloads last month
- 169
Model tree for OrionLLM/GRM-2.5-Air
Collection including OrionLLM/GRM-2.5-Air
Evaluation results
- Diamond on Idavidrein/gpqa View evaluation results leaderboard 12.5
- Mmlu Pro on TIGER-Lab/MMLU-Pro View evaluation results 43.6