Aetheris Playground
A hybrid Mamba-MoE multilingual model — 800M parameters distilled from tiny-aya-global (3.35B) with 4.2x compression. Chat in any of 67 languages and it responds in kind.
Polyglot Conversation
Chat with Aetheris in any language — it detects your language and responds naturally. Try switching languages mid-conversation to see multilingual capabilities.
Try These
Click any example to load it:
Full Language Coverage
67 Languages Across 6 Regions
Aetheris inherits multilingual coverage from CohereLabs/tiny-aya-global, spanning 13 language families, 15 scripts, and languages from every inhabited continent.
Model Comparison
Teacher vs Student Architecture
Organization | GQA (16 query / 4 kv heads) | SSM (Mamba) + MoE routing |
Math Reasoning
mGSM — Multilingual Grade School Math
Multi-step arithmetic reasoning across languages. 8-shot evaluation.
Causal Reasoning
XCOPA — Cross-lingual Causal Reasoning
Choice of plausible alternatives — tests commonsense and causal reasoning.
Compression Quality
Quality Retention Across Languages
Percentage of teacher performance preserved at 4.2x compression. Target: >80%.
Performance
Inference Throughput
Batch size 1, sequence length 256. Student achieves ~3x speedup with 4x memory reduction.
Multilingual Equity
Degradation Equity Score
DES measures fairness of quality loss across language families. 0 = perfectly equitable, 1 = maximum inequity. Red = needs remediation.
High-risk languages (targeted for extra training): sw, te, ar, am, my, lo, gu, km
Training
Training Progress
Key metrics tracked across the 3-stage distillation pipeline.