Darmm/darmm-chat-kazakh-8b

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Darmm/darmm-chat-kazakh-8b is an 8 billion parameter conversational assistant developed by Darmm, fine-tuned from Qwen3-8B. It specializes in the Kazakh language, while retaining Russian and English capabilities, and is optimized for chat-based interactions. This model significantly improves performance on Kazakh-specific benchmarks like Belebele and KazMMLU compared to its base model. It is best suited for applications requiring conversational AI in Kazakh.

Loading preview...

Overview

Darmm/darmm-chat-kazakh-8b is an 8 billion parameter conversational AI model, fine-tuned from Qwen/Qwen3-8B. Its primary focus is on providing a robust chat assistant experience in Kazakh, while also maintaining proficiency in Russian and English. The model was developed by Darmm through QLoRA fine-tuning on approximately 105,000 instruction pairs, comprising 65,000 Kazakh, 25,000 Russian, and 15,000 English examples.

Key Capabilities and Performance

This model demonstrates notable improvements in Kazakh language understanding and generation:

  • Enhanced Kazakh Performance: Achieves 39.8% on the Belebele (kaz_Cyrl) benchmark and 38.3% on KazMMLU (Kazakh subjects), significantly outperforming the base Qwen3-8B model (27.2% and 27.9% respectively).
  • Multilingual Support: While specialized for Kazakh, it retains conversational abilities in Russian and English.
  • Optimized for Chat: Designed as a conversational assistant, superseding previous text generation models.

Usage Recommendations

  • Sampling over Greedy Decoding: To prevent degeneration into repetition, it is crucial to use sampling settings (e.g., do_sample=True, temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.1) for generation.
  • Non-Thinking Mode: The model is trained for enable_thinking=False in chat templates.

Limitations

  • Factual Reliability: The model has limited factual reliability (38% KazMMLU) and should not be used as a sole source for factual information without verification.
  • Language Nuances: Due to machine-translated training data, Kazakh phrasing can occasionally be unnatural.
  • Decoding Sensitivity: Greedy decoding leads to degraded output; sampling is mandatory for optimal results.

Finetunes
2 models