AFS (AI Foundry Service)

Low Cost, Low Threshold, High Efficiency, High Security
One-stop Generative AI Solution to Help Enterprises Build Exclusive Large Models.

What is AFS?

The innovative model of Taiwan’s semiconductor industry, particularly in wafer foundry services, has driven the global IC design sector and the overall semiconductor and technology industry ecosystem to decades of vigorous growth. Now, generative AI is igniting a productivity revolution and signaling the arrival of the era of Moore’s Law for artificial intelligence systems!

Taiwan AI Cloud launched AFS (AI Foundry Service) to help enterprises efficiently build exclusive large language models through AIHPC high-speed computing power: In the AI ​​model training (optimization) stage, it provides the “Formosa Foundation Model, FFM” with Traditional Chinese enhancement. Enterprises just need to prepare data before starting large model training, which effectively saves enterprises time and cost for the initial setup; in the AI ​​model inference (deployment) stage, in addition to “cloud” deployment services, it also provides the market’s only integrated hardware and software “on-premises” deployment solution.

From ChatGPT to IndustrialGPT, AFS is the enterprise’s AI foundry to help efficiently build dedicated models that align with corporate culture and real needs, models that are unique, and practical which can ensure high-security requirements for enterprise data security, compliance, and privacy, and can be safely connected to the enterprise’s internal systems. With the advantages of low cost and low threshold, it is specially designed to empower enterprises’ generative AI applications and deployment.

AFS Features and Advantages

From Optimization to Deployment
With Confidentiality & Exclusive Large Models that ChatGPT Cannot Achieve

Pre-trained “FFM” integration with 176 Billion Parameters.
A solid foundation for enterprises to build applications without starting from scratch.

Low Cost

A supercomputer environment with pay-as-you-go billing for you to scale according to your usage, from just thousands to tens of thousands of dollars, in which you can quickly deploy enterprise-grade language model optimization services.

Help enterprises effectively control costs, avoid waste, and save huge setup costs, reducing development risks and investment in hardware equipment and manpower.

Low Threshold

FFM-BLOOMZ and FFM-LLama2 traditional Chinese LLM and cloud and local model deployment services enable enterprises to build models without starting from scratch, also, the capabilities in traditional Chinese semantic understanding and text generation are more practical.

Help enterprises quickly get started develop their own large models, and choose appropriate solutions based on their respective scales and needs.

High Efficiency

Based on AIHPC high-speed computing power and a no-code platform, optimization of 100 million tokens for enterprises can be completed in as little as 6 hours, while a single 176B training can handle up to 2.1 billion tokens.

Assist enterprises to accelerate the training and deployment of exclusive large models, save optimization costs, and enhance enterprise efficiency.

High Security

It meets the requirements for enterprise cybersecurity, SLA, maintenance services, compliance and auditability, can be commercially authorized, adopts a tenant isolation system, and provides the only deployment solution in the market that meets various local cybersecurity needs.

Effectively ensures the confidentiality of sensitive data and record privacy for enterprise users while reducing cybersecurity risks and providing secure protection.

AFS One-stop LLMs Solution

Cost Savings Nearly 0 Times

As a comparison, building your own AIHPC system and pre-training large models would cost billions.
Assist enterprises in optimizing and deploying models in a high-efficiency, high-security environment.


Optimization Solution

AFS Platform

Provide enterprises with the “lowest cost” and “highest efficiency” large language model training environment.

\ Limited Availability!/

On-premises Solution

AFS Appliance

Provide enterprises with the ability to directly deploy their own large language models in local data centers/private clouds while ensuring both cybersecurity and privacy.


Optimized LLM
Hosting Solution

AFS Cloud

Offer flexible deployment solutions for enterprises, pay as you go to eliminate unnecessary costs.


Native LLM
Hosting Solution

AFS ModelSpace

Provide enterprises with a variety of open source large language models and flexibility of allocating computing resources.

AFS Application Cases

Start building enterprise-specific large language models with AFS.

External Data
Enterprise Knowledge Management
Numerical Predictions
for the Future
Application Cases
Inference Service

AI Foundry Service x AIHPC Computing Power
On-premises LLM Solution
Limited Availability ‧ Apply Now


The innovative model of semiconductor foundries in Taiwan has driven the global IC design industry and the overall semiconductor and technology industry to thrive for several decades. Taiwan AI Cloud introduces AI Foundry Service to provide industrial services for computing power. It also offers opportunities for every enterprise to develop its strengths based on the open computing resources provided by Taiwan AI Cloud and aims to establish an AI 2.0 industry ecosystem of co-creation and mutual prosperity. By launching new business models and services in the AI service industry, it is driving the transformation of Taiwan’s industries with AI.

If you have model optimization needs, you can refer to the following scenarios for selection:

Scenario 1: For enterprises that require model training based on actual task needs and want to pay based on actual computational resource usage (Pay-As-You-Go), you can subscribe to the AFS Platform immediately.

If you have model deployment and inference needs, you can refer to the following scenarios for selection:

Scenario 2: For customers with sensitive data and security considerations, you can choose the AFS Appliance, which allows you to deploy the enterprise’s fine-tuned large language models on-premises.

Scenario 3: For customers with cloud-based needs, you can choose the AFS Cloud to perform model inference services through cloud hosting. You will be billed based on the GPU hours used.

Taiwan AI Cloud’s AFS soluitons provide enterprise users with dedicated “enterprise-level” generative AI solutions that meet the requirements of data security, SLA, operational services, and compliance according to specific needs. AFS also provides complete one-stop services, including POC validation for enterprise users, customized model optimization (fine-tuning), and deployment and inference of models in hybrid mode, both on-premises and in the cloud. Users can choose the appropriate service based on their specific needs. Please rest assured that Taiwan AI Cloud will not retain any customer data or records from the use of AFS services or any interaction with other products, nor will such data be used as FFM training data.

The entire range of AFS services is built in compliance with national cybersecurity regulations. Your service and data are hosted in non-governmental A-grade computing facilities and enjoy the highest level of information security. For even higher sensitivity requirements, you can choose to deploy dedicated models on-premises using “AFS Appliance”
(On-premises deployment solution for large language models) provided by Taiwan AI Cloud. This ensures that internal enterprise data remains within the premises and eliminates the risk of data leakage.

If the amount of internal and external data in a specific enterprise-specific domain reaches the order of 1 billion tokens, it is recommended to use Pretrain-SFT two-stage optimization training to establish the best-performing enterprise-specific domain model. You can also add a new dictionary based on BLOOM 250K tokenizer to allow patent translation and exclusive glossaries to be included in the enterprise’s exclusive domain knowledge. For example, If a specific term must always correspond to a particular translation term or have a domain-specific meaning, creating a dedicated dictionary will be as effective as incorporating the new dictionary during the pretraining stage.

If there is insufficient data for pretraining, you can also directly perform SFT. However, it is recommended to have a dataset calculated by 100,000s for effective SFT, such as with an example range of 100,000 to 500,000.

Scenario 1: {“inputs” (prompt/question): “Please help me translate to Chinese\n\n how are you?” “targets” (answer): “你好嗎?”}. The “\n” in the prompt sentence is a newline symbol.

Data cleaning instructions:

(1) At the end of the training input prompt (Q), you can use a new line to separate the prompt/context. There is no need to use special separators (for example: “\n\n###\n\n”)

(2) The data cannot contain line breaks (\r), invisible characters (e.g. tab \t), or special characters (e.g. \ /). Special characters must be removed or escape characters must be added or converted to specific text before training.

(3) Input/output data may include punctuation marks at the end but is not required.

(4) When there are duplicates in the training input prompt (Q), it is recommended to manually remove them first.

(5) Enterprises must first de-identify the data (all kinds of sensitive information within the exclusive domain of the enterprise) before uploading it to the TWCC cloud for model pre-training or fine-tuning. Additionally, enterprises can also use the second-stage SFT for de-identification task training, and train the LLM to perform de-identification of content based on conditions defined by the enterprise.

(6) After completing de-identification SFT, you can perform content de-identification tasks based on the prompt. Before performing SFT de-identification tasks, please prepare training data for various de-identification scenarios based on industry, compliance, or preferences for replacement characters.

The format is a jsonl file, where each sample data contains an “inputs” and “targets” pair, and the total number of tokens for each data entry cannot exceed 2048.

FFM is a model developed by Taiwan AI Cloud based on the Large Language Model (LLM) technology, with an impressive parameter count of up to 176 billion. The original version of the model is BLOOM open source, which allows for commercial licensing and usage. Taiwan AI Cloud’s in-house technology research and development team, based on years of experience in the NLP field, has enhanced the performance of the model by strengthening the LLM technique of the pre-trained model. It demonstrates relatively significant-high-quality performance in the understanding of Traditional Chinese language semantics and knowledge domain. While specifically targeting the diverse application needs of Taiwanese enterprise users in various domains, the FFM also demonstrates high-quality results in text content generation and has the capability that encompasses global knowledge and multiple languages.

Yes, it can support. LangChain simplifies the development of applications using large language models. By leveraging the customization capabilities provided by LangChain (refer to the following link:, application developers can write custom LLM wrappers based on the documentation and specify the address for deploying the FFM large language model within the wrapper.

We provide ready-to-use Container Compute Service (CCS) for Generative AI with a variety of AI Framework and open-source model choices, such as FFM-BLOOM, Embedding, Meta Llama 2, and also support user-installation of Stable Diffusion on CCS to meet various AI-generated text and images technology development needs.

Create a account to enjoy 60 days of experience for 10,000 credits for free now! Sign up here.


「AFS (AI Foundry Service)」提供一站式整合服務,協助企業開發專屬的企業級生成式AI大模型


Free Consultation Service

Contact our experts to learn more and get started with the best solution for you.

Apply for AFS

AFS (AI Foundry Service)


Business Inquiries/ Distribution Cooperation