Hon Hai Research Institute Launches Traditional Chinese LLM With Reasoning Capabilities
Hon Hai Research Institute has launched FoxBrain, the first Traditional Chinese Large Language Model (LLM), developed in just four weeks using 120 NVIDIA H100 GPUs. Based on Meta Llama 3.1 architecture with 70B parameters, FoxBrain demonstrates superior performance in mathematics and logical reasoning compared to existing models.
The model, backed by Foxconn (TWSE:2317), will be open-sourced and was initially designed for internal applications including data analysis, decision support, and code generation. FoxBrain utilized efficient training methods and resource optimization, processing 98B tokens of high-quality pre-training data across 24 topic categories.
The model's development showcases Taiwan's AI capabilities, achieving results comparable to world-class AI models despite computational resources. FoxBrain will be integrated into Foxconn's three major platforms: Smart Manufacturing, Smart EV, and Smart City, with its results to be presented at NVIDIA GTC 2025 on March 20.
Il Hon Hai Research Institute ha lanciato FoxBrain, il primo Modello di Linguaggio di Grandi Dimensioni (LLM) in cinese tradizionale, sviluppato in soli quattro settimane utilizzando 120 GPU NVIDIA H100. Basato sull'architettura Meta Llama 3.1 con 70 miliardi di parametri, FoxBrain dimostra prestazioni superiori in matematica e ragionamento logico rispetto ai modelli esistenti.
Il modello, supportato da Foxconn (TWSE:2317), sar脿 reso open-source ed 猫 stato inizialmente progettato per applicazioni interne, tra cui analisi dei dati, supporto decisionale e generazione di codice. FoxBrain ha utilizzato metodi di addestramento efficienti e ottimizzazione delle risorse, elaborando 98 miliardi di token di dati di pre-addestramento di alta qualit脿 suddivisi in 24 categorie tematiche.
Lo sviluppo del modello mette in mostra le capacit脿 dell'IA di Taiwan, raggiungendo risultati comparabili a modelli di IA di livello mondiale nonostante le limitazioni delle risorse computazionali. FoxBrain sar脿 integrato nelle tre principali piattaforme di Foxconn: Smart Manufacturing, Smart EV e Smart City, con i risultati che saranno presentati al NVIDIA GTC 2025 il 20 marzo.
El Instituto de Investigaci贸n Hon Hai ha lanzado FoxBrain, el primer Modelo de Lenguaje de Gran Tama帽o (LLM) en chino tradicional, desarrollado en solo cuatro semanas utilizando 120 GPU NVIDIA H100. Basado en la arquitectura Meta Llama 3.1 con 70 mil millones de par谩metros, FoxBrain demuestra un rendimiento superior en matem谩ticas y razonamiento l贸gico en comparaci贸n con los modelos existentes.
El modelo, respaldado por Foxconn (TWSE:2317), ser谩 de c贸digo abierto y fue dise帽ado inicialmente para aplicaciones internas, incluyendo an谩lisis de datos, soporte en la toma de decisiones y generaci贸n de c贸digo. FoxBrain utiliz贸 m茅todos de entrenamiento eficientes y optimizaci贸n de recursos, procesando 98 mil millones de tokens de datos de preentrenamiento de alta calidad en 24 categor铆as tem谩ticas.
El desarrollo del modelo muestra las capacidades de IA de Taiw谩n, logrando resultados comparables a modelos de IA de clase mundial a pesar de las limitaciones de recursos computacionales. FoxBrain se integrar谩 en las tres principales plataformas de Foxconn: Smart Manufacturing, Smart EV y Smart City, con los resultados que se presentar谩n en el NVIDIA GTC 2025 el 20 de marzo.
頇嶍晿鞚� 鞐瓣惮靻�臧赌 韽姢敫岆爤鞚�鞚� 於滌嫓頄堨姷雼堧嫟. 鞚措姅 鞝勴喌 欷戧淡鞏措 霅� 觳� 氩堨Ц 雽�攴滊 鞏胳柎 氇嵏(LLM)搿�, 120臧滌潣 NVIDIA H100 GPU毳� 靷毄頃橃棳 雼� 4欤� 毵岇棎 臧滊皽霅橃棃鞀惦媹雼�. 70鞏� 臧滌潣 毵り皽氤�靾橂ゼ 臧赌歆� 氅旐儉 霛茧 3.1 鞎勴偆韰嶌矘毳� 旮半皹鞙茧 頃橂┌, 韽姢敫岆爤鞚胳潃 旮办〈 氇嵏鞐� 牍勴暣 靾橅暀 氚� 雲茧Μ鞝� 於旊鞐愳劀 鞖办垬頃� 靹彪姤鞚� 氤挫棳欷嶋媹雼�.
鞚� 氇嵏鞚� 韽姢旖� (TWSE:2317)鞚� 歆�鞗愳潉 氚涭晞 鞓ろ攬 靻岇姢頇旊悹 鞓堨爼鞚措┌, 雿办澊韯� 攵勳劃, 鞚橃偓 瓴办爼 歆�鞗� 氚� 旖旊摐 靸濎劚 霌� 雮措秬 鞎犿攲毽紑鞚挫厴鞚� 鞙勴暣 觳橃潓 靹り硠霅橃棃鞀惦媹雼�. 韽姢敫岆爤鞚胳潃 須湪鞝侅澑 頉堧牗 氚╇矔瓿� 鞛愳洂 斓滌爜頇旊ゼ 頇滌毄頃橃棳 24臧� 欤检牅 旃错厡瓿犽Μ鞐愳劀 98鞏� 臧滌潣 瓿犿拡歆� 靷爠 頉堧牗 雿办澊韯半ゼ 觳橂Μ頄堨姷雼堧嫟.
氇嵏鞚� 臧滊皽鞚� 雽�毵岇潣 AI 鞐焿鞚� 氤挫棳欤茧┌, 旎错摠韺� 鞛愳洂鞐愲弰 攵堦惮頃橁碃 靹戈硠鞝� 靾橃鞚� AI 氇嵏瓿� 牍勱祼頃� 毵岉暅 瓴瓣臣毳� 雼劚頄堨姷雼堧嫟. 韽姢敫岆爤鞚胳潃 韽姢旖橃潣 靹� 臧赌歆� 欤检殧 頂岆灚韽检澑 鞀る韸� 鞝滌“, 鞀る韸� 鞝勱赴彀�, 鞀る韸� 鞁滍嫲鞐� 韱淀暕霅� 鞓堨爼鞚措┌, 攴� 瓴瓣臣電� 2025雲� 3鞗� 20鞚� NVIDIA GTC鞐愳劀 氚滍憸霅� 瓴冹瀰雼堧嫟.
Le Hon Hai Research Institute a lanc茅 FoxBrain, le premier Mod猫le de Langage de Grande Taille (LLM) en chinois traditionnel, d茅velopp茅 en seulement quatre semaines avec 120 GPU NVIDIA H100. Bas茅 sur l'architecture Meta Llama 3.1 avec 70 milliards de param猫tres, FoxBrain d茅montre des performances sup茅rieures en math茅matiques et en raisonnement logique par rapport aux mod猫les existants.
Le mod猫le, soutenu par Foxconn (TWSE:2317), sera open-source et a 茅t茅 initialement con莽u pour des applications internes, y compris l'analyse de donn茅es, le soutien 脿 la d茅cision et la g茅n茅ration de code. FoxBrain a utilis茅 des m茅thodes d'entra卯nement efficaces et l'optimisation des ressources, traitant 98 milliards de tokens de donn茅es de pr茅-entra卯nement de haute qualit茅 r茅parties sur 24 cat茅gories th茅matiques.
Le d茅veloppement du mod猫le met en avant les capacit茅s en IA de Ta茂wan, atteignant des r茅sultats comparables 脿 ceux des mod猫les d'IA de classe mondiale malgr茅 des ressources informatiques limit茅es. FoxBrain sera int茅gr茅 dans les trois principales plateformes de Foxconn : Smart Manufacturing, Smart EV et Smart City, avec ses r茅sultats pr茅sent茅s lors du NVIDIA GTC 2025 le 20 mars.
Das Hon Hai Research Institute hat FoxBrain gestartet, das erste gro脽e Sprachmodell (LLM) in traditionellem Chinesisch, das in nur vier Wochen mit 120 NVIDIA H100 GPUs entwickelt wurde. Basierend auf der Architektur Meta Llama 3.1 mit 70 Milliarden Parametern zeigt FoxBrain 眉berlegene Leistungen in Mathematik und logischem Denken im Vergleich zu bestehenden Modellen.
Das Modell, unterst眉tzt von Foxconn (TWSE:2317), wird als Open Source ver枚ffentlicht und wurde urspr眉nglich f眉r interne Anwendungen wie Datenanalyse, Entscheidungsunterst眉tzung und Code-Generierung entwickelt. FoxBrain nutzte effiziente Trainingsmethoden und Ressourcenoptimierung und verarbeitete 98 Milliarden Token hochwertiger Vortrainingsdaten in 24 Themenkategorien.
Die Entwicklung des Modells zeigt die KI-F盲higkeiten Taiwans und erzielt Ergebnisse, die mit weltklasse KI-Modellen vergleichbar sind, trotz begrenzter Rechenressourcen. FoxBrain wird in die drei Hauptplattformen von Foxconn integriert: Smart Manufacturing, Smart EV und Smart City, und die Ergebnisse werden am 20. M盲rz 2025 auf der NVIDIA GTC pr盲sentiert.
- Development of proprietary AI technology demonstrating competitive capabilities
- Efficient resource utilization: 4-week development vs industry standards
- Superior performance metrics in mathematics and reasoning compared to existing models
- Strategic application across three major business platforms
- Still trails DeepSeek's distillation model in mathematical reasoning performance
First version by AI Research Center performs well in mathematics and reasoning
The institute, which is backed by Hon Hai Technology Group ("Foxconn") (TWSE:2317), the world's largest electronics manufacturer and leading technological solutions provider, said the LLM 鈥� code named FoxBrain 鈥� will be open sourced and shared publicly in the future. It was originally designed for applications used in the Group's internal systems, covering functions such as data analysis, decision support, document collaboration, mathematics, reasoning and problem solving, and code generation.
FoxBrain not only demonstrates powerful comprehension and reasoning capabilities but is also optimized for Taiwanese users' language style, showing excellent performance in mathematical and logical reasoning tests.
"In recent months, the deepening of reasoning capabilities and the efficient use of GPUs have gradually become the mainstream development in the field of AI. Our FoxBrain model adopted a very efficient training strategy, focusing on optimizing the training process rather than blindly accumulating computing power," said Dr. Yung-Hui Li, Director of the Artificial Intelligence Research Center at Hon Hai Research Institute. "Through carefully designed training methods and resource optimization, we have successfully built a local AI model with powerful reasoning capabilities."
The FoxBrain training process was powered by 120 , scaled with InfiniBand networking, and finished in just about four weeks. Compared with inference models recently launched in the market, the more efficient and lower-cost model training method sets a new milestone for the development of
FoxBrain is based on the Meta Llama 3.1 architecture with 70B parameters. In most categories among TMMLU+ test dataset, it outperforms Llama-3-
- Established data augmentation methods and quality assessment for 24 topic categories through proprietary technology, generating 98B tokens of high-quality pre-training data for Traditional Chinese
- Context window length: 128 K tokens
- Utilized 120聽NVIDIA H100 GPUs for training, with total computational cost of 2,688 GPU days
- Employed multi-node parallel training architecture to ensure high performance and stability
- Used a unique Adaptive Reasoning Reflection technique to train the model in autonomous reasoning
In test results, FoxBrain showed comprehensive improvements in mathematics compared to the base Meta Llama 3.1 model. It achieved significant progress in mathematical tests compared to Taiwan Llama, currently the best Traditional Chinese large model, and surpassed Meta's current models of the same class in mathematical reasoning ability. While there is still a slight gap with DeepSeek's distillation model, its performance is already very close to world-leading standards.
FoxBrain's development 鈥� from data collection, cleaning and augmentation, to Continual Pre-Training, Supervised Finetuning, RLAIF, and Adaptive Reasoning Reflection 鈥� was accomplished step by step through independent research, ultimately achieving benefits approaching world-class AI models despite limited computational resources. This large language model research demonstrates that
Although FoxBrain was originally designed for internal group applications, in the future, the Group will continue to collaborate with technology partners to expand FoxBrain's applications, share its open-source information, and promote AI in manufacturing, supply chain management, and intelligent decision-making.
During model training, NVIDIA provided support through the
The results of FoxBrain is scheduled to be shared for the first time at a major conference during NVIDIA GTC 2025 Session Talk "" on March 20.
About Hon Hai Research Institute
The institute has five research centers. Each center has an average of 40 high technology R&D professionals, all of whom are focused on the research and development of new technologies, the strengthening of Foxconn's technology and product innovation pipeline, efforts to support the Group's transformation from "brawn" to "brains", and the enhancement of the competitiveness of Foxconn's "3+3" strategy.
About Foxconn .
View original content to download multimedia:
SOURCE Hon Hai Research Institute
FAQ
What are the key technical specifications of FoxBrain, Hon Hai's new LLM (HNHPF)?
How does FoxBrain's performance compare to other Traditional Chinese language models?
When will Hon Hai (HNHPF) make FoxBrain available as open-source?
What business applications will FoxBrain serve for Hon Hai (HNHPF)?
