Tag: Formosa-1

APMIC announced today (April 29th) that its lightweight 3B-parameter language model "Formosa-1," developed in collaboration with the Twinkle AI community, can now be deeply integrated with the Model Context Protocol (MCP). Formosa-1 is Taiwan's first 3B-parameter Traditional Chinese language model that can run on mobile devices. Through the MCP protocol, Formosa-1 can connect to internal and external enterprise systems, enabling natural language-driven information retrieval, process operation, and decision support, creating a new intelligent human-computer interaction interface. It also supports integration with external APIs, ERP systems, and IoT devices. Combined with the fine-tuning and distillation solutions provided by APMIC's S1 model outsourcing service, it helps enterprises build their own edge models. Whether deployed on low-to-mid-range GPUs or embedded platforms, it can run smoothly and respond instantly, significantly reducing the technical and cost barriers to implementing AI solutions. This meets the needs of industrial equipment manufacturers, technology companies, mobile device brands, and robotics and HMI application fields for edge deployment, low power consumption, and rapid response. Natural Language Moves from "Communication" to "Control": Unlike traditional language models that are primarily for chatting, "Formosa-1" incorporates the MCP protocol, enabling deep integration with internal enterprise systems. Through language-driven mechanisms, the model not only responds to queries but also serves as a smart command interface to trigger business processes, data retrieval, and decision support, further expanding the practical applications of natural language. This solution has demonstrated its effectiveness in diverse scenarios. For example, in industrial automation, engineers can use voice to query CNC error codes or activate specific modules; retailers can use voice to operate POS systems and quickly obtain sales data and inventory information; in logistics control, the language model can instantly check inventory status and assign delivery tasks to specific drivers, significantly improving operational efficiency. Leveraging the MCP protocol, "Formosa-1" can flexibly connect to external APIs such as ERP systems, weather data, exchange rate information, and machine monitoring, allowing users to trigger complex operational processes through natural language, replacing traditional interfaces or manual input and enhancing interactive intuitiveness. APMIC's "Formosa-1" can now be deeply integrated with enterprise internal systems via the MCP protocol, comprehensively expanding the application boundaries of language models. APMIC founder and CEO Po-Han Wu stated, "Promoting the low-cost deployment of AI models across diverse devices will be a key path to achieving AI popularization. APMIC's self-developed 3B model combined with the MCP protocol is an important step in moving human-computer interaction from a 'cloud dream' to 'embedded reality.' APMIC has long been committed to developing localized language models and enterprise private AI deployment solutions. The 'Formosa-1' Traditional Chinese model demonstrates APMIC's strength in model refinement and on-the-ground deployment. The current integration of the 3B model and the MCP protocol shows that APMIC is once again driving AI technology into more fields through practical applications, enabling more enterprises, venues, and devices to truly realize the ability to drive business actions in real time through language commands." The "Formosa-1" Traditional Chinese model connects to the exchange rate information MCP protocol, allowing users to understand real-time exchange rates through commands. With its lightweight deployment architecture, Formosa-1 enables real-time execution of language models on the ground. Designed with 3B parameters, it offers high deployment flexibility, running smoothly on mobile phones, IoT devices, robots, human-machine interfaces (HMIs), and various edge devices without relying on cloud computing resources or high-end GPUs, significantly lowering the barrier to entry. This feature allows enterprises to flexibly deploy language models on mobile devices and remote locations, expanding intelligent applications to diverse scenarios such as manufacturing, retail, and logistics. While not specifically designed for deep logical reasoning, Formosa-1 demonstrates superior performance in handling structured queries, automating task processes, and integrating multi-modal data. To assist enterprises in customizing models to meet industry needs, APMIC has also launched the "Embedded Business Model Fine-Tuning Solution." This solution can tailor model versions with industry-specific language and command understanding capabilities based on internal knowledge bases and workflows, further enhancing the accuracy and value of the model in practical applications.

APMIC (Accelerate Private Machine Intelligence Company), a Taiwanese-founded AI solutions provider and an ecosystem partner of NVIDIA's fine-tuning models, announced a collaboration with Twinkle AI, a Traditional Chinese language model research community, to launch "Formosa-1," Taiwan's first mobile-compatible Traditional Chinese inference model with 3 billion parameters. APMIC also partnered with Twinkle AI to open-source the efficient evaluation framework "Twinkle Eval," used to assess AI model performance and promote the development and local application of AI technology in Taiwan. "Formosa-1," developed by APMIC and the Twinkle AI community with technical support and experience sharing from the National Center for High-Performance Computing (NCHC) research team, is Taiwan's first large-scale language model with 3 billion parameters that can run on mobile devices. This model was trained using APMIC PrivAI's product model built on the NVIDIA NeMo end-to-end platform, employing distillation techniques, and its model weights are fully open under the MIT license, thereby promoting the development of open-source Traditional Chinese AI technology applications in Taiwan. To enhance its model's reasoning capabilities, Formosa-1 is trained using data aligned with the Taiwan Chain of Thought (TCOT) framework and leverages NVIDIA NeMo Data Curator to accelerate data management, ensuring superior performance in legal reasoning, logical deduction, and mathematical reasoning. Regarding language data construction, Formosa-1's training data encompasses 100 billion data points and includes high-quality Traditional Chinese tokens from diverse texts such as news, legal documents, academic papers, and social discussions, ensuring accurate understanding and application of Traditional Chinese contexts by the AI. The open-source evaluation framework "Twinkle Eval," designed for large-scale reasoning models, is integrated with NVIDIA NeMo...