What challenges did Azercell Telecom LLC face before AI implementation?

Adapt foundation models to a morphologically rich language with limited training data and no existing blueprint for efficient LLM training in Azerbaijani. Build a production-ready framework that could scale to larger corpora and larger models.

What was the impact of the AI solution?

23% higher training throughput. 58% lower peak GPU memory usage. 2x improvement in tokens per word and a production-ready scalable framework for independent operation.

Source

Azercell trains an Azerbaijani LLM using Amazon SageMaker AI (FSDP + LoRA)

Use case typeAI model trainingUpdated May 28, 2026

Azercell Telecom LLC built an Azerbaijani large language model on Amazon SageMaker AI for telecom use cases and a customer-facing chatbot using a custom tokenizer, continued pre-training, and LoRA fine-tuning. The implementation used distributed training with PyTorch FSDP and Liger Kernel optimizations, plus Amazon SageMaker Unified Studio, Amazon EC2, Amazon S3, and Amazon CloudWatch.

Organization: Azercell Telecom LLC
Industry: Tech & Comms
Location: Azerbaijan
Published: May 2026

Reported outcomes

−58%

peak GPU memory usageOther quantified impact

+23%training throughput−50%tokens per word+100%content capacity in context window+700%max batch size

Strategic outcomes

Scale & capacityBuilt a production-ready scalable frameworkCustomer experience & trustEnabled a customer-facing chatbotEmployee experienceAllowed Azercell to operate the framework independentlyOther strategic outcomeTurned a base model into a conversational assistant

Primary read

Use case focus

Showing 2 of 2

1AI model training
2Conversational assistants

Adapt foundation models to a morphologically rich language with limited training data and no existing blueprint for efficient LLM training in Azerbaijani.
Build a production-ready framework that could scale to larger corpora and larger models.

Train a custom Azerbaijani tokenizer to reduce token fragmentation.
Continue pre-training Llama 3.2 1B on Amazon SageMaker AI with FSDP and Liger Kernel optimizations.
Apply LoRA fine-tuning to transform the model into a conversational assistant.

Technologies

Amazon SageMaker AI Amazon SageMaker Unified Studio Amazon EC2 Amazon S3 Amazon CloudWatch PyTorch Fully Sharded Data Parallel Liger Kernel LoRA Hugging Face Transformers Hugging Face tokenizers Llama 3.2 1B

23% higher training throughput.
58% lower peak GPU memory usage.
2x improvement in tokens per word and a production-ready scalable framework for independent operation.

Architecture

A three-stage training pipeline: (1) custom Azerbaijani tokenizer development with BBPE and a 100k vocabulary, (2) continued pre-training of Llama 3.2 1B on Amazon SageMaker AI training jobs using PyTorch FSDP and Liger Kernels, and (3) LoRA supervised fine-tuning to turn the model into a conversational assistant. Training jobs were launched from Amazon SageMaker Unified Studio, with artifacts in Amazon S3 and metrics tracked in TensorBoard and Amazon CloudWatch.

Implementation partners1

AWS Generative AI Innovation Center

Sources & evidence1

Groundedness: 5/5Type: Blog PostPublished: May 28, 2026Publisher: AWSEvidence: VendorConfidence: Medium

Primary source

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases

Customer evidence