AksaraLLM/aksarallm-1.5b-v2-checkpoint

TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kPublished:Apr 16, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

AksaraLLM/aksarallm-1.5b-v2-checkpoint is an earlier 1.5 billion parameter Qwen2-based causal language model developed by AksaraLLM, specifically trained for Indonesian language tasks. This checkpoint offers a context length of 32768 tokens and is primarily maintained for historical reference and reproducibility. It has been superseded by a newer version with improved perplexity and reduced English language leakage, making it less suitable for new Indonesian-focused applications.

Loading preview...

AksaraLLM/aksarallm-1.5b-v2-checkpoint Overview

This model, developed by AksaraLLM, is an earlier 1.5 billion parameter checkpoint based on the Qwen2 architecture, primarily focused on the Indonesian language. It features a substantial context length of 32768 tokens.

Key Characteristics

  • Architecture: Qwen2ForCausalLM with 1777.1 million parameters.
  • Language Focus: Primarily Indonesian, though it exhibits a higher English-stopword ratio (3.6%) compared to its successor.
  • Performance Baseline: Achieves a perplexity of 9.9 on short Indonesian sentences, as measured by a Devin audit.

Status and Usage

This v2-checkpoint is maintained for historical reference and reproducibility purposes. For any new development or downstream applications, AksaraLLM explicitly recommends using the newer AksaraLLM/AksaraLLM-Qwen-1.5B-v5-public model. The v5-public version demonstrates measurably better perplexity (8.4) and a significantly lower English leak (0.9% English-stopword ratio), making it a more robust choice for current Indonesian language tasks.

License

The model is released under the Apache 2.0 License.