# Large Language Models (LLMs)

Large Language Models (LLMs) are neural networks trained to understand and generate human language at scale.

At Dhee, we work **exclusively with transformer-based LLMs**, as they currently provide the most reliable foundation for high-quality, multilingual, and conversational systems.

Transformer architectures enable models to process entire sequences in context, making them especially suitable for complex language understanding and generation tasks.

<figure><img src="https://2566128011-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FUn1oKrw7P8P5qiTjWXjv%2Fuploads%2FH11ETk7bCtfmjyxeepCq%2Fthe-annotated-transformer_14_0.png?alt=media&#x26;token=da45f7eb-3a78-4abe-be52-e19f3fbdab48" alt=""><figcaption></figcaption></figure>

### Transformer-Based LLMs

Transformer-based LLMs model language by learning relationships between all tokens in a sequence simultaneously.

This allows them to:

* Maintain long-range context
* Capture nuanced grammatical and semantic relationships
* Scale efficiently with data and model size

These properties make transformers the dominant architecture behind modern LLMs used in production systems today.

### What Transformer-Based LLMs Are Good At

Transformer-based LLMs excel at:

* Natural language understanding
* Conversational response generation
* Translation across languages
* Summarization and rewriting
* Intent recognition
* Semantic similarity and entailment

They perform best when:

* Sufficient context is provided
* The task is language-centric
* High linguistic fidelity is required

### How Transformer-Based LLMs Work (Conceptually)

At a high level, transformer-based LLMs operate as follows:

1. **Tokenization**\
   Text is converted into tokens suitable for the model.
2. **Contextual Processing**\
   The transformer processes all tokens together, allowing each token to attend to every other token in the context.
3. **Token Prediction**\
   The model predicts the most likely next token, repeatedly, to generate an output.

This process enables coherent, context-aware language generation.

### Stateless by Design

Transformer-based LLMs are **stateless**.

Each interaction:

* Is processed independently
* Has no inherent memory of previous turns
* Relies entirely on provided context

Any persistent behavior are implemented **outside the model in Dhee GPT Platform.**

### Conversational LLMs in Indian Languages

A core focus at Dhee is building **conversational transformer-based LLMs for Indian languages**.

These models are designed for:

* Native conversational flow
* Spoken-language patterns
* Code-mixed inputs
* Multi-turn dialogue consistency

Rather than treating Indian languages as translation targets, these models are trained and adapted for **direct conversational competence**.

Our open model collection is available here for you to try and use in your projects:\
👉 <https://huggingface.co/collections/dheeyantra/dhee-nxtgen-qwen3-v2>