ecloudtech/Erk-32B

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Erk-32B is a 32 billion parameter Turkish language model developed by eCloud Tech., built upon the Qwen3-32B base model. It underwent 1.05 billion tokens of Turkish continued pre-training and 423 million tokens of Turkish instruction tuning, resulting in significant improvements in Turkish language understanding and generation. The model excels in Turkish morphological tasks and general Turkish language capabilities, making it suitable for Turkish-specific NLP applications.

Loading preview...

Erk-32B: A Specialized Turkish Language Model

Erk-32B, developed by eCloud Tech., is a 32 billion parameter Turkish language model based on the open-source Qwen3-32B architecture. It has been extensively continued-pretrained on 1.05 billion Turkish tokens and instruction-tuned with 423 million Turkish tokens (373k cleaned examples) to enhance its performance in the Turkish language.

Key Capabilities and Performance

  • Enhanced Turkish Language Understanding: Erk-32B v2 achieves 71.5% on TurkishMMLU (clean 652 items), a notable improvement over the base Qwen3-32B's 67.64%. It is the highest-scoring model on TurkishMMLU among 17 measured models.
  • Superior Morphological Generation: On TurkMorfBench v2 generation, Erk-32B scores 59.38% with the system prompt, significantly outperforming the base model's 31.64%.
  • Robust Evaluation Protocol: All reported metrics are rigorously validated through contamination audits, option rotation, forced-choice measurements, and paired bootstrap confidence intervals, ensuring reliability.
  • Identity via System Prompt: The model's identity is provided through a default system prompt in the chat template, explicitly stating its origin and base model.

Use Cases and Recommendations

Erk-32B v2 is recommended for general-purpose Turkish assistant applications. Its strong performance in Turkish language tasks, particularly in morphological generation and overall understanding, makes it a suitable choice for developers building Turkish-centric NLP solutions. The model is available in merged weights and GGUF quantizations (Q8_0, Q4_K_M) for various deployment needs.

Finetunes
2 models