π FastVLA: The "Arabic Hero" Model (7B)
I trained a 7B-parameter Robot to understand Arabic for $0.48/hr. Stop renting H100s.
FastVLA democratizes Vision-Language-Action (VLA) models by enabling real-time robotics reasoning on budget hardware (NVIDIA L4/T4). This model is fine-tuned specifically for the Arabic language, bridging the gap in localized Physical AI.
π Performance Benchmarks (Tesla T4)
| Metric | OpenVLA (Base) | FastVLA (Fine) | Improvement |
|---|---|---|---|
| Inference Latency | 1420.0 ms | 198.2 ms | 7.16x faster |
| Action Error (L2) | 28.5 px | 12.4 px | 2.30x more accurate |
| VRAM Usage | 5.50 GB | 4.45 GB | 19.2% reduction |
π Why this matters
In 2026, 81% of Arabic AI is still text-only. FastVLA is the first bridge to Arabic Physical AI, allowing robotics policies to understand native instructions while running at 5Hz control loops on standard cloud instances.
π₯ Usage
from fastvla import FastVLAModel
model = FastVLAModel.from_pretrained(
"hamzabouajila/fastvla-arabic-hero",
load_in_4bit=True
)
# Arabic Command: "Push the block to the target"
instruction = "Ψ―ΩΨΉ Ψ§ΩΨΨ¬Ψ± Ψ₯ΩΩ Ψ§ΩΩΨ―Ω"
action = model.predict_action(image, instruction)
π§ͺ Training Details
- Architecture: OpenVLA-7B + Triton Continuous Action Head
- Optimization: Unsloth 4-bit QLoRA
- Dataset: hamzabouajila/ar-pusht-image
- Hardware: Trained on 2x NVIDIA T4 ($0.80/hr total)
Maintained by: Hamza Bouajila | GitHub Repository
- Downloads last month
- 3