Gnani Evon v3.3: a 30B hybrid MoE model for Indic languages

Gnani released gnani-evon-v3.3-30B-A3B, a Mamba2-Transformer hybrid MoE model with 30B total and ~3.5B active parameters and a 128K context. It leads open-weight comparators on Indic extractive QA (XQuAD-in 64.59 F1) but trails Gemma and Sarvam on several other benchmarks.

• 30B total parameters, ~3.5B active per token, 131,072-token context
• XQuAD-in 64.59 F1 and XorQA-in 41.83 F1 lead open-weight comparators
• MILU macro mean 78.74 across 11 languages including Hindi and Tamil
• Trails Gemma-4 and Sarvam on CrossSum-in, IndicIFEval and BharatMath