Skip to main content
Back

Rta-AI Labs' Nandi-Mini Encodes Indian Languages 7x Faster Than Qwen on Edge Devices

A 150M parameter transformer trained from scratch on 525B tokens delivers English and 10 Indic languages for edge deployments.

Pro

Model, Full Access
  • Rta-AI Labs released Nandi-Mini-150M, a 150M parameter multilingual model under Apache 2.0.
  • Pre-trained from scratch on 525B tokens covering English and 10 Indic languages.
  • Uses factorized embeddings and layer sharing (16x2 layers, effective depth 32) for efficiency.
  • Beats SmolLM and SmolLM2 at the same size on MMLU, GPQA, GSM8K, and HumanEval.
  • Tokenization fertility on Indic scripts is 5-8x better than Qwen3 and SmolLM3.
  • Base, Instruct, and tool-calling variants available now.

A small research outfit called Rta-AI Labs (also publishing under FrontiersMind) has released Nandi-Mini-150M, a compact multilingual language model aimed squarely at edge devices and on-prem workloads that need to handle Indian languages. It ships under Apache 2.0 in both base and instruct variants, with 500M and 1B siblings already in training.

The pitch is to maximize performance per parameter through architectural efficiency rather than scale, optimizing for edge devices, on-prem deployments, and low-latency applications. What makes it worth a look is how it handles Indic scripts, where most Western tokenizers fall apart.

Under the hood

Nandi-Mini-150M is a causal decoder pre-trained from scratch on 525 billion tokens. The architecture combines the usual modern ingredients (RoPE positional embeddings, RMSNorm, SwiGLU activations, grouped-query attention) with two efficiency tricks that matter more at this scale:

  • Factorized embeddings shrink the parameter cost of a large vocabulary. The vocab is 131,072 tokens, huge for a 150M model, but factorization keeps the memory footprint manageable.
  • Layer sharing reuses 16 physical layers to give an effective depth of 32. You get the representational benefit of a deeper stack without paying twice for weights.