ApX logoApX logo

Hunyuan Lite

Active Parameters

7B

Context Length

250K

Modality

Text

Architecture

Mixture of Experts (MoE)

License

Tencent Hunyuan Community License

Release Date

30 Oct 2024

Knowledge Cutoff

Aug 2024

System Requirements

VRAM requirements for different quantization methods and context sizes

1,024 tokens

16.34 GB VRAM

Consumer

1x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

250,000 tokens

50.61 GB VRAM

Consumer

3x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

Architecture Diagram

Input TokensToken EmbeddingPosition: AbsoluteHidden: 4.1k · Context: 250Kx 32 layersRMSNormPre-AttentionMulti-Head Attention32Q / 8KV headsHead dim: 128+RMSNormPre-FFNSparse MoE FFNSwiGLU+Final RMSNormOutput Logits

Evaluation Benchmarks

No evaluation benchmarks for Hunyuan Lite available.

Rankings

Overall Rank

-

Coding Rank

-

About Hunyuan Lite

Hunyuan Lite is a specialized, text-based language model developed by Tencent, engineered to deliver sophisticated linguistic and reasoning capabilities within a compact computational footprint. Part of the broader Hunyuan ecosystem, this model is designed for deployment on edge devices such as laptops, smartphones, and in-vehicle systems. Its primary objective is to provide a highly efficient solution for natural language understanding, code generation, and complex mathematical problem-solving without the high resource overhead typically associated with large-scale models. By optimizing the balance between performance and latency, the model enables advanced AI integration in environments where memory and power consumption are critical constraints.

The architectural framework of the 7B variant employs a dense Transformer-based structure, departing from the Mixture of Experts (MoE) design used in its larger counterparts like Hunyuan-Large or Hunyuan-A13B. A defining technical innovation of this series is its support for an ultra-long context window of 256,000 tokens, which allows for the ingestion and analysis of extensive documents, complete books, or lengthy conversation histories. The model integrates Grouped Query Attention (GQA) to accelerate inference speed and reduce the memory footprint of the KV cache. Additionally, it features a unique dual-mode reasoning capability, enabling users to switch between a "fast-thinking" mode for immediate responses and a "slow-thinking" mode that utilizes chain-of-thought processing for deeper analytical tasks.

Hunyuan Lite is optimized for versatile deployment and is compatible with mainstream inference frameworks like vLLM, SGLang, and TensorRT-LLM. The model adopts a Rotary Position Embedding (RoPE) scheme to maintain stability across its expanded context window and utilizes SwiGLU activation for enhanced expressive power in its feed-forward layers. Engineered for agentic workflows, it demonstrates high proficiency in tool-use and structured data generation. The release of open weights under a community license facilitates specialized fine-tuning and integration into private-domain knowledge engines and automated assistant platforms.

Technical Specifications

Attention

Attention Structure

Multi-Head Attention

Attention Heads

32

Key-Value Heads

8

Attention Head Dimension

-

Position Embedding

Absolute Position Embedding

RoPE Theta

-

Sliding Window Attention

-

Sliding Window Size

-

Sliding Window Ratio

-

Linear Attention

-

Linear Attention Ratio

-

Normalization

RMS Normalization

Activation Function

SwigLU

Dimensions

Hidden Dimension Size

4,096

Number of Layers

32

FFN Intermediate Size (Dense)

-

Multi-Token Prediction Heads

-

Tokenizer

Vocabulary Size

-

Mixture of Experts

Total Expert Parameters

-

Number of Experts

-

Active Experts

-

Shared Experts

-

FFN Intermediate Size (per Expert)

-

Dense Layers Before MoE

-

About Hunyuan

Tencent Hunyuan large language models with various capabilities.


Other Hunyuan Models