ceyda/Qwen3-0.6B-Base-trim-koen-32768

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 31, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ceyda/Qwen3-0.6B-Base-trim-koen-32768 is a 0.8 billion parameter Qwen3-based causal language model, derived from Qwen/Qwen3-0.6B-Base. This version features a vocabulary trimmed to 32,768 tokens, specifically optimized for Korean and English, resulting in a 20.5% reduction in parameters. It maintains numerical identity with the original model for kept tokens and is designed for efficient processing of Korean and English prose.

Loading preview...

Model Overview

This model, ceyda/Qwen3-0.6B-Base-trim-koen-32768, is a specialized version of the Qwen/Qwen3-0.6B-Base model. Its primary distinction lies in its vocabulary trimming, which reduces the byte-level BPE vocabulary from approximately 151.7k to 32,768 tokens. This trimming focuses on Korean and English languages, including all 256 byte-level tokens, 26 special tokens, and the most frequent Korean/English tokens along with their full BPE merge-derivation closure.

Key Characteristics

  • Parameter Reduction: The trimming process reduces the model's parameters from 596 million to 474 million, a 20.5% decrease, making it more efficient.
  • Numerical Equivalence: Despite the vocabulary reduction, the model's weights for the kept tokens are copied verbatim, ensuring numerical identity with the original Qwen3-0.6B-Base model on these tokens.
  • Lossless Encoding: Verified to provide lossless round-trip encoding for Korean, English, and code. Greedy generation on natural Korean/English text is token-for-token identical to the base model.

Intended Use Cases

  • Korean and English Prose: Specifically optimized for processing and generating text in Korean and English.
  • Efficiency: Benefits from reduced parameter count and vocabulary size for potentially faster inference or lower memory footprint in relevant applications.

Limitations

  • This is a base model and is not instruction-tuned.
  • While code and rare jargon encode losslessly, they may result in slightly longer token sequences compared to the original model.
  • Text in languages other than Korean and English will fall back to byte tokens, leading to longer sequences.

Finetunes
1 models