Typhoon-2-70B
Parameters
70B
Context Length
128K
Modality
Text
Architecture
Dense
License
Apache-2.0
Release Date
1 Jun 2024
Knowledge Cutoff
Dec 2023
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
7x RTX 4090
24GB VRAM
Datacenter
2x NVIDIA A100
80GB VRAM
Apple Silicon
2x Apple M3 Max
128GB VRAM
128,000 tokens
Consumer
10x RTX 4090
24GB VRAM
Datacenter
3x NVIDIA A100
80GB VRAM
Apple Silicon
2x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
No evaluation benchmarks for Typhoon-2-70B available.
Rankings
Overall Rank
-
Coding Rank
-
About Typhoon-2-70B
Typhoon-2-70B is a high-capacity Thai-English large language model developed by SCB 10X, specifically architected to address the linguistic complexities of the Thai language. Built upon the Llama 3.1 70B backbone, this model undergoes extensive continual pre-training on a curated corpus of over 5 billion high-quality Thai tokens. This training process is designed to align the model with Thai cultural nuances and linguistic structures while preserving the original English reasoning capabilities of the underlying architecture. The resulting model serves as a foundation for enterprise-level applications requiring high precision in bilingual contexts.
The technical architecture employs a dense, decoder-only transformer structure with Grouped-Query Attention (GQA) to optimize inference efficiency and memory throughput. It utilizes a 128K token context window, enabling the processing of lengthy legal documents, technical manuals, and multi-turn conversational histories. The model integrates advanced post-training techniques, including supervised fine-tuning (SFT) and Direct Preference Optimization (DPO), to enhance its instruction-following accuracy and function-calling capabilities. These optimizations allow the model to interact with external tools and APIs, facilitating complex agentic workflows.
Released under the Llama 3.1 Community License, Typhoon-2-70B provides a transparent path for developers to integrate sovereign AI capabilities into production environments. Its design emphasizes performance in specialized Thai domains such as legal reasoning, cultural content generation, and sophisticated data analysis. By bridging the gap between English-centric foundation models and local language requirements, Typhoon-2-70B enables the development of localized AI solutions that maintain parity with global standards of reasoning and accuracy.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
64
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
8,192
Number of Layers
80
FFN Intermediate Size (Dense)
-
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
-
About Typhoon
Typhoon is a Thai language model family developed by SCB 10X. It is specifically optimized for the Thai language, addressing complexities such as the lack of word delimiters and tonal nuances. The models are trained on Thai-centric datasets including legal, cultural, and historical documents to ensure localized context and knowledge.