YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
GRM-3-Nano
Base Architecture: Qwen 3.5 0.8B (natively multimodal, early-fusion transformer) License: Apache 2.0 Context Window: 262K tokens (extensible to 1M) Modality: Text, Image, Video
NOTICE- GRM-3-Nano is currently under final beta testing evaulation, and will release in the next 2-3 Weeks!
Overview
GRM-3-Nano is a compact, natively multimodal model built on Qwen 3.5 0.8B, engineered for fast, private, on-device inference. It unifies text, vision, and coding capabilities in a single sub-1B-parameter footprint, avoiding bolt-on vision adapters through early-fusion multimodal training.
Key Features
- Runs fully offline on laptops, phones, and legacy hardware with no internet dependency
- Native multimodal token training for unified text/image/video understanding
- Strong OCR and document layout comprehension
- 201 languages supported natively
- Extensible context window up to 1M tokens for large documents and codebases
On-Device Reasoning
GRM-3-Nano outperforms models several times its size on reasoning and vision benchmarks despite its small parameter count. Its early-fusion architecture achieves cross-generational parity with larger Qwen3 models, outperforming Qwen3-VL on reasoning, coding, agent, and visual understanding tasks.
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| MMLU (general knowledge/reasoning) | 42.3 | Strong for sub-1B scale |
| OCRBench (document/image-text vision) | 79.1 | Robust document and image reasoning |
| Overall score vs. flagship 397B model | ~54% | Baseline among 0.8B-9B model family |
| Context window | 262K (up to 1M) | Supports PDFs, large codebases |
| Language support | 201 languages | Native multilingual coverage |
Ideal Use Cases
- Edge and embedded deployment with limited or no GPU access
- Lightweight agents and mobile assistants
- Offline document and image analysis
- Applications prioritizing low resource footprint over maximum reasoning depth
Notes
NexLM is open sourcing this model as a fantastic, sub 1B param, reasoning and on device vision model.
Training Factors
Locally trained on M4 Pro, 24GB RAM Macbook Pro using Unsloth Studio. (15 days Pretrain, 10 days eval) Datasets used (NexLM open source and proprietary):
- CompReasoning (NexLM)- Open Source [HF]
- SciencePhilosphy (NexLM)- Open Source[HF]
- Soul Doc/ What is Identity (NexLM)- NexLM Proprietary
- Deep reasoning/Probing questions
- Safety/Red-teaming (NexLM)- NexLM Proprietary
- How to Code (NexLM)- NexLM Proprietary
- CodeSonnet5 (NexLM)- Open Source [HF]
- ComplexMath (NexLM)- NexLM Proprietary
- Grammer/English (NexLM)- NexLM Proprietary
- Identity/GRM-3-Nano Identity docs (NexLM)- NexLM Proprietary
- Anthropic/hh-rlhf (Anthropic) - Open Source [HF]
- openai/frontierscience (OpenAI)- Open Source [HF]
- Tool calling/usage (NexLM)- NexLM Proprietary
- Post training- response formatting
- Post training- Safety/Red-teaming (NexLM)- NexLM Proprietary
NOTICE- GRM-3-Nano is currently under final beta testing evaulation, and will release in the next 2-3 Weeks!