Efficient enterprise inference
Planned for lower-cost deployment profiles and faster inference across practical enterprise workloads.
Model Architecture
Decoder-only transformer architecture for multilingual production inference.
Tassili is planned as a transformer-based model family optimized for multilingual tokenization, distributed training, and enterprise-ready deployment profiles.
Planned Model Sizes
A staged model family with different enterprise tradeoffs.
Balanced reasoning and cost
Positioned as the most balanced variant for multilingual reasoning and operational affordability.
Advanced reasoning
Planned for deeper contextual understanding and more demanding reasoning tasks.
Future roadmap: MoE (Mixture-of-Experts) experimental variant.
Training Strategy
A structured pipeline from pre-training to domain adaptation.
Pre-training
- Curated multilingual corpus
- Balanced French / Arabic / English / Darija token distribution
- Deduplication and filtering
- Quality-based dataset weighting
Estimated training scale target: multi-trillion token regime, scaled progressively.
Instruction tuning
- Enterprise use cases
- Administrative writing
- Technical Q&A
- Multilingual dialogue
Alignment & safety
- Human feedback loops
- Bias mitigation layers
- Content moderation safeguards
- Domain-sensitive output constraints
Domain adaptation layer
- Finance DSLM
- Energy DSLM
- Legal DSLM
- Education DSLM
- Telecom DSLM
Tokenization Strategy
Optimized for Arabic morphology and French syntax.
Tassili uses a multilingual tokenizer designed to better capture Arabic morphological complexity while preserving French syntactic structure and English technical precision.
The architecture is built to support multilingual reasoning, domain adaptation, and production deployment from the outset.