Taiwan’s Ministry of Digital Affairs, through the Administration for Digital Industries, is in the final stretch of applications for its year 115 free AI computing platform. The last application round closes on Oct. 30.
The subsidy is aimed at eligible domestic businesses. Once approved, applicants can use the platform’s GPU computing resources free of charge for three months, and they may apply multiple times based on demand. Companies can use their own industry data to train or fine-tune open-source models, or connect directly through APIs to preinstalled large language models on the platform to build application services.
Application schedule has entered the later stage
In a news release, the Administration for Digital Industries said that by the end of year 114, a total of 186 AI startups and information service providers had already used the platform, producing at least 266 models or innovative applications.
Applications for year 115 have now moved into the later rounds. For training, only the second and third batches remain, with deadlines on Aug. 31 and Oct. 30. For inference, the remaining windows are the fourth, fifth, and sixth batches, with deadlines on Aug. 31, Sept. 24, and Oct. 30. The full-year application period ends on Oct. 30, and late applicants will need to wait until the next annual cycle.
How the program works and who can apply
Applications are submitted entirely online, and businesses must verify their identity with a business certificate before filing. Reviews follow a rolling intake model with batch-based review, rather than a fixed submission cycle, so companies can apply according to their own research and development timeline.
Eligibility is limited to five categories of registered companies or business entities: e-commerce and mail-order businesses, software publishing, computer programming and related services, data processing and information supply services, and other financial support services that include third-party payments.
Sole proprietorships, partnerships, and limited partnerships are all eligible organizational forms. Startups established after year 107 and operating for less than eight years receive priority status.
Training modes and follow-up support
The application portal is hosted on the platform’s official website. Training applications are split into two paths. Under the container-based environment, companies package and upload their own models and training tools, while the platform handles execution. Under the no-code training environment, users do not need to write code. They can select built-in large language models on the platform, upload their own data, and begin training a domain-specific model.
According to the official description, approved teams with strong performance may also receive priority access to the “Enhanced Investment in AI Startups Implementation Plan,” which includes follow-up support such as fundraising matchmaking.
Models available on the platform
On the training side, locally developed options include the National Science and Technology Council’s TAIDE 8B model, the Taiwan academia-industry alliance’s TAME Llama-3-Taiwan-70B, Hon Hai Research Institute’s FoxBrain-70B, and the FFM series developed by the operating unit, TaiSmart Cloud.
International models available for training include Llama 3.1, Llama 3.3, Microsoft’s Phi-4-reasoning series, Mistral’s Magistral-Small, and NVIDIA’s Nemotron-Super-49B, which was tuned from Llama 3.3 70B. For image recognition, the platform supports YOLO. For speech recognition, it supports Whisper.
The inference category adds another group of models that can only be connected through APIs and cannot be directly fine-tuned with model weights. That list includes openai/gpt-oss-120b, Meta’s full Llama 4 lineup including Maverick and Scout, and Gemma 4 31B. For speech recognition, the platform also adds Taiwan Tongues ASR, which supports Taiwanese and other languages used in Taiwan.
Still behind the cutting edge
The report noted that the free models still trail the most advanced systems currently available. The newest model on the list is Gemma 4, which Google open-sourced on April 2 this year. As for gpt-oss-120b, OpenAI’s own positioning describes its core reasoning performance as close to o4-mini, a mid-tier reasoning model, leaving a noticeable gap versus flagship models from the same period.
Whether the program is worth applying for depends on the applicant’s needs. The report said the offer may be a practical fit for teams that need vertical-domain fine-tuning, face data localization or compliance limits that prevent sending data out, or run enough inference volume that lowering cost becomes a priority. In those cases, free GPU access together with open-source weights that can be fine-tuned may prove useful.