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:Transformer 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-specific tasks, demonstrating a +3.37 gain on TurkishMMLU and a +27.73 gain on TurkMorfBench v2 generation compared to its base.

Loading preview...

Erk-32B: A Specialized Turkish Language Model

Erk-32B is a 32-billion parameter language model developed by eCloud Tech., based on the open-source Qwen3-32B foundation. It has been extensively adapted for the Turkish language through 1.05 billion tokens of continued pre-training and 423 million tokens of instruction tuning (373k cleaned examples).

Key Capabilities and Performance

This model demonstrates notable improvements in Turkish language tasks, with all reported metrics rigorously validated through a comprehensive measurement protocol that includes contamination audits, option rotation, forced-choice measurements, and paired bootstrap confidence intervals.

  • TurkishMMLU: Achieves 71.01% on a clean subset of TurkishMMLU (652 items), a +3.37 point increase over the base Qwen3-32B model (67.64%).
  • TurkMorfBench v2 Generation: Scores 59.38% with a system prompt, marking a substantial +27.73 point gain compared to the base model's 31.64%. This indicates strong performance in Turkish morphological generation.
  • Morphological Knowledge: While generation improved significantly, the gain in core morphological knowledge (forced-choice tasks) over the base model was not demonstrably significant (+1.99 with a confidence interval including zero).
  • Identity: The model's identity is provided via a default system prompt in the chat template, explicitly stating its origin and base model.

Training and Technical Details

Erk-32B was trained using LoRA (r=256, α=512) for continued pre-training and two rounds of instruction tuning. The training involved approximately 37,000 core-hours on 4x A100-80GB GPUs. The final release is a merged weight at a LoRA scale of 0.75. GGUF quantizations (Q8_0, Q4_K_M) are also available for efficient deployment.

Good for

  • Applications requiring high-quality Turkish language understanding and generation.
  • Tasks benefiting from improved Turkish morphological accuracy in generation.
  • Developers seeking a specialized Turkish LLM built on a robust open-source foundation.

Finetunes
2 models