Accuracy vs Scale in Arabic & English
TorchLM plotted against leading models on MMLU, showing strong accuracy per parameter for multilingual problem solving
MMLU Benchmark Comparison
Arabic MMLU
English MMLU
TorchLM / ALLaM
Modern
Legacy
Arabic Acc
3D Model Explorer
Each sphere represents an AI model positioned in 3D space:
- → Right = higher English accuracy
- ↑ Up = higher Arabic accuracy
- ↗ Forward = more parameters
Larger spheres indicate better efficiency (accuracy per billion parameters).
TorchLM-8b achieves top bilingual performance at a fraction of the size.
Drag to rotate, scroll to zoom, hover for details.
TorchLM
Arabic Acc
→= efficiency
Exceptional efficiency by design.
More accuracy per parameter than models 10× its size.
TorchLM-8b delivers top-tier bilingual performance while remaining compact enough to deploy anywhere—from cloud to on-device. Smaller footprint means faster inference, lower costs, and greener AI.
COMPACT
FAST
GREEN
Efficiency Leaderboard
Accuracy per billion parameters (higher = more efficient)
1
TorchLM-8b
97.52
ALLaM-7B
90.73
Qwen2-7B
85.74
Mistral-7B
75.75
Meta-Llama-3-8B
74.46
gemma-7B-it
65.77
Aya-23-8B
65.08
AceGPT-7B
65.09
Yi-9B
64.410
Llama-2-7B-Chat
57.9