CraneAILabs/edu-ganda-gemma-e2b-v5

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The CraneAILabs/edu-ganda-gemma-e2b-v5 is a 5.1 billion parameter Gemma-4-E2B model developed by Crane AI Labs, specifically designed as an offline Luganda assistant for Primary 1-3 teachers. This model excels at lesson planning, phonics drills, assessment management, and parent communication, optimized to run efficiently on mobile devices. It is a SLERP merge of a pedagogy-tuned model and a dictionary-trained language model, enhancing both instructional capabilities and Luganda vocabulary. The model features a 32768 token context length and is fine-tuned for exact numeracy and translation tasks.

Loading preview...

EduGanda-Gemma-4-E2B-v5: Offline Luganda Teacher Assistant

This model, developed by Crane AI Labs, is the culmination of their Proposal 2 work, funded by the Gates Foundation. It's a 5.1 billion parameter Gemma-4-E2B variant, specifically engineered to function as an offline Luganda assistant for Primary 1-3 teachers, designed for mobile phone deployment.

Key Capabilities & Features

  • Teacher Support: Plans lessons, drills phonics, sets and marks assessments, and drafts messages to parents.
  • Hybrid Architecture: A SLERP merge of two Crane Gemma-4-E2B checkpoints:
    • v5-Polished: Focuses on pedagogy fine-tuning, reinforcement learning for exact numeracy, and translation.
    • v4: Contributes Luganda vocabulary and instruction-style training from its dictionary-trained language line.
  • Multilingual Focus: Demonstrates strong performance in Luganda translation (chrF++ scores of 48.4 Luganda→English and 41.9 English→Luganda).
  • Numeracy: Achieves 72.0% on 150 held-out numeracy problems.
  • Instruction Adherence: Scores 72.3% on 220 program-checked prompts.
  • Optimized for Mobile: Sized and designed to run efficiently on phone devices.

Limitations & Usage Notes

  • Luganda Content: May still exhibit errors in Luganda content; review before use with learners. Judged Luganda quality is 2.67 out of 10.
  • Conciseness: Follows system prompts closely; request complete artifacts rather than concise answers.
  • Repetition Guard: Recommended serving settings include repetition_penalty=1.0 with no_repeat_ngram_size=4 to prevent loops.
  • Stop Tokens: Uses Gemma-4's <turn|> (token 106) as an end-of-turn token; do not use Gemma-3's <end_of_turn>.

Finetunes
1 models