AI Wiki

Explore AI Wiki

Citation and evidence

FoxBrain

16 min full readUpdated 20 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision history

Topics

AI ModelsLarge Language ModelsReasoning Models

Cite this article

FoxBrain is a family of large language models developed by the AI Research Center of Hon Hai Research Institute (HHRI, 鴻海研究院), the research arm of Foxconn (Hon Hai Technology Group). HHRI announced the first version on 10 March 2025 as a Traditional Chinese reasoning model built on Meta's Llama 3.1 70B and trained on 120 NVIDIA H100 GPUs in about four weeks.[1][3] Version 1.2 was published on Hugging Face on 31 July 2025 as gated weights for academic and research use.[7][8] Foxconn uses FoxBrain as the base model for internal tools and for its three platforms (smart manufacturing, smart electric vehicles and smart cities). In 2026 the company began extending it into multimodal and physical AI models, starting with FoxBrain 2.0, which is based on NVIDIA Nemotron.[11][12][13] An NVIDIA case study published in October 2026 describes FoxBrain as the core model of Foxconn's "AI factory," trained on an NVIDIA DGX SuperPOD built from DGX GB200 systems.[14][15]

Key facts

AttributeDetail
DeveloperAI Research Center, Hon Hai Research Institute (Foxconn)[1]
First announced10 March 2025[1]
Base model (v1.x)Meta Llama 3.1 70B; Hugging Face lists Llama-3.1-70B-Instruct as the fine-tuning base[1][6]
Parameters (v1.x)70B (70,553,706,496 BF16 parameters in the published checkpoint)[1][6]
Context window128K tokens (v1.0 announcement)[1]
LanguagesTraditional Chinese (Taiwanese usage) and English[1][6]
Public releaseLlama_3.1-FoxBrain-70B (preview, May 2025); Llama_3.1-FoxBrain-70B-V1.2 (31 July 2025)[6][7][8]
Access terms (v1.2)Gated; academic institutions and research organizations only, no commercial use without authorization[7]
Later versionFoxBrain 2.0, announced on 1 June 2026 for introduction at NVIDIA GTC Taipei, built on NVIDIA Nemotron Omni[11][12]

Expanded article table

Background

Hon Hai Research Institute is backed by Foxconn. According to its own description, it has five research centers, each with about 40 R&D staff, and it supports the group's "3+3" strategy and its move from "brawn" to "brains."[1] Speaking at an AI forum in May 2025, Dr. Yung-Hui Li (栗永徽), director of HHRI's AI Research Center, said Foxconn had worked for more than 20 months after the release of ChatGPT to produce its first-generation model.[5] He gave four reasons for training an in-house model instead of relying on foreign ones: protecting core technology, building unique domain data, training AI talent, and controlling key AI technology. He also cited the limited Traditional Chinese and local-culture understanding of international models, the security risk of sending sensitive data to foreign providers, and the difficulty of customizing models controlled by others.[5] He said Foxconn's plan was to become a platform provider: a base model would sit between its hardware and its three platform services.[5]

In a November 2025 interview, Li described the shortage of private domain data as the main weakness of companies such as OpenAI, Google and Anthropic. Foxconn, by contrast, has access to data from its electronics production lines, its EV business and its partner hospitals.[10] The FoxBrain_LLMs GitHub repository linked from the model cards was created in July 2023. Its citation block credits "Tran Nhiem, HHRAI."[20]

FoxBrain 1.0 (March 2025)

HHRI announced FoxBrain on 10 March 2025. The release headline called it HHRI's first Traditional Chinese LLM "with reasoning capabilities," and the release said it was optimized for the language style of Taiwanese users.[1][2] Reuters reported Foxconn's statement that it was "Taiwan's first large language model with reasoning capabilities that is optimised for traditional Chinese and Taiwanese language styles."[3] The model was originally designed for Foxconn's internal systems. Its listed functions were data analysis, decision support, document collaboration, mathematics, reasoning and problem solving, and code generation.[1]

Specifications and training

ItemValue (as stated by HHRI)
Base architectureMeta Llama 3.1, 70B parameters[1]
Pre-training data98B tokens of Traditional Chinese pre-training data, generated with proprietary data augmentation and quality assessment across 24 topic categories[1]
Context window128K tokens[1]
Training hardware120 NVIDIA H100 GPUs, scaled with NVIDIA Quantum-2 InfiniBand networking[1]
Total compute2,688 GPU-days[1]
Training timeAbout four weeks[1][3]
Training frameworkNVIDIA NeMo, with support from NVIDIA's Taipei-1 supercomputer and technical consultation[1][4]
ParallelismMulti-node parallel training architecture[1]

Expanded article table

HHRI described the pipeline in stages: data collection, cleaning and augmentation, continual pre-training, supervised fine-tuning, RLAIF (reinforcement learning from AI feedback), and a proprietary method it calls "Adaptive Reasoning Reflection," which it said trains the model to reason autonomously.[1] Li said the team followed an "AI training AI" approach: because usable Traditional Chinese data was limited at first, it built its own data pipeline for cleaning, augmentation and evaluation to generate Taiwan-specific training data with minimal manual work.[5] In his launch statement, Li said the team had focused "on optimizing the training process rather than blindly accumulating computing power."[1]

Reuters described Taipei-1 as the largest supercomputer in Taiwan, owned and operated by NVIDIA in Kaohsiung.[3]

Reported performance

All published FoxBrain results are self-reported by Foxconn. As of 28 October 2026, no independent third-party evaluation had been located.

  • TMMLU+. HHRI said FoxBrain outperformed Llama-3-Taiwan-70B in most TMMLU+ categories, with its biggest gains in mathematics and logical reasoning. It also said FoxBrain improved on the base Llama 3.1 model across mathematics.[1] TMMLU+ is a Traditional Chinese multiple-choice benchmark covering 66 subjects, released in 2024 as a larger successor to TMMLU.[18] Llama-3-Taiwan-70B comes from the Taiwan LLM project, a line of Llama models adapted to Taiwanese Mandarin.[19]
  • Comparison with DeepSeek. Foxconn said a "slight gap" remained between FoxBrain and DeepSeek's distillation model, but that its performance was "very close to world-leading standards."[1][3]
  • MT-Bench. The v1.0 model card reports MT-Bench comparisons against Meta-Llama-3.1-70B-Instruct and Taiwan-LLaMA-70B-Instruct using pairwise comparison, voting preference and overall rating (human and LLM-judge). It says FoxBrain "demonstrated favorable results in pairwise human judgment" but gives no scores in the text.[6]
  • News rewriting. The v1.0 card lists a 100-sample Traditional Chinese news-rewriting test in which FoxBrain scored 68.50, against 62.26 for "Claude" and 61.19 for "ChatGPT." It does not say which model versions were tested.[6]

HHRI was scheduled to present FoxBrain's results for the first time at the NVIDIA GTC 2025 session "From Open Source to Frontier AI: Build, Customize, and Extend Foundation Models" on 20 March 2025.[1]

Hugging Face release

The March 2025 announcement said FoxBrain would be open-sourced at a later date.[1] In May 2025 Li said HHRI planned to publish 8B and 70B models with technical documentation on Hugging Face by the end of May. He also said Foxconn would show at Computex how FoxBrain could be integrated into NVIDIA NIM microservices.[5] He described future 1.x versions as adding tool calling and distilled small models, with a 2.0 version moving toward about 100B parameters and multimodality.[5]

Two repositories appeared under the FoxconnAI organization on Hugging Face:

RepositoryCreatedNotes
FoxconnAI/Llama_3.1-FoxBrain-70B19 May 2025Version 1.0. The card calls it a preview with access "temporarily limited to collaborating partners" and gives the planned official release as late June to early July[6]
FoxconnAI/Llama_3.1-FoxBrain-70B-V1.231 July 2025Version 1.2, with "dual thinking modes"[7]

Expanded article table

In a talk at the AI WAVE event, Li announced that FoxBrain had been open-sourced on Hugging Face on 31 July 2025 for academic and research use. He said the released version was 1.2, because the team had worked through the night to fix issues found in internal testing before posting it.[8] Both repositories are gated. The v1.2 page asks applicants to email Foxconn before submitting an access request. Its usage agreement grants access "exclusively to academic institutions and research organizations," prohibits commercial use without explicit authorization, and says commercial availability would come later through authorized channels such as AWS under a separate license.[7] The agreement also requires compliance with the Llama 3.1 license and Taiwanese law, and lists prohibited uses including military applications, heavy industry control and critical infrastructure operations.[7] As of 28 October 2026, the Hugging Face listing showed only these two FoxBrain repositories in the FoxconnAI organization, both 70B Llama 3.1 derivatives.[6][7]

Version 1.2 thinking modes

The v1.2 model card describes two reasoning modes, selected with a thinking_mode parameter in the chat template:[7]

ModeBehavior described on the model card
Budget_Thinking (default)Allocates a reasoning "budget" of 1 to 9 steps based on problem complexity and outputs tagged steps, reflections and quality scores
Extend_ThinkingReasons freely inside <think></think> tags; the card says a presence_penalty of 1.5 is critical for reliable output

Expanded article table

Custom system prompts take priority over both modes. The card reports improvements of 15 percent in multi-step reasoning (Budget_Thinking) and 22 percent in complex analytical tasks (Extend_Thinking), and lists them in the reasoning analysis that follows its Taiwan MMLU+ evaluation note, but does not name a baseline or give underlying scores.[7] For inference it recommends vLLM on 2 to 8 H100 GPUs, with 4 recommended.[7] The card's version history lists 1.0 (Chinese language proficiency), 1.1 (better reasoning and efficiency) and 1.2 (dual thinking modes), and names 2.0 as the version that would add industrial domain knowledge.[7]

Applications at Foxconn

Foxconn says FoxBrain is meant to power its three platforms: Smart Manufacturing, Smart EV and Smart City.[1] In November 2025, Li said FoxBrain was still at the language-model stage and that internal uses were mostly text-based: coding support, paper analysis and equipment repair instructions. If a machine on a production line failed, for example, an employee could ask FoxBrain to explain the inspection steps and run simple checks without waiting for a specialist.[10]

FoxBrain is also the language model in Nurabot, a nursing-assistant robot built by Foxconn and Kawasaki Heavy Industries and piloted at Taichung Veterans General Hospital. NVIDIA described the robot in May 2025 as combining the FoxBrain LLM, virtual training with Isaac for Healthcare, and onboard compute on the NVIDIA Holoscan sensor processing platform running on an NVIDIA Jetson Orin device. Foxconn estimates the robot can cut nurses' workload by up to 30 percent when used for medication delivery, specimen transport and ward patrols.[9] The same NVIDIA post described FoxBrain as developed with the NeMo framework, trained on NVIDIA Hopper GPUs, and capable of text-to-speech, automatic speech recognition and natural language processing.[9]

At Hon Hai Tech Day 2025, Foxconn chairman Young Liu said in his opening remarks that FoxBrain can understand Confucian thought.[10] Li said the team expects specialized versions for manufacturing, smart healthcare and electric vehicles as the base model improves.[10]

NVIDIA's October 2026 case study describes a tiered process. Foxconn starts from open foundation models (Llama, Gemma, NVIDIA Nemotron and others), adds proprietary data to build the base FoxBrain LLM, then runs further pre-training and fine-tuning on industrial, urban and healthcare data. This produces two specialized platforms: Genesis Platform for smart manufacturing workflows and CityGPT for urban management and citizen services.[14] Foxconn fine-tunes FoxBrain for manufacturing, vision and robotics tasks with NeMo and serves the specialized platforms with NVIDIA Dynamo for low-latency inference.[14] The case study also mentions a FoxBrain vision-language model built by continuing training from the FoxBrain LLM.[14] Foxconn's subsidiary Visionbay.ai, an NVIDIA Cloud Partner, offers the stack as AI-as-a-service to governments and enterprises in Asia.[14][15]

Roadmap and FoxBrain 2.0

In November 2025, Li laid out a four-stage roadmap:[10]

StagePlanned capability
FoxBrain 1.0Text processing (LLM)
FoxBrain 2.0Adds vision as a vision language model
FoxBrain 3.0A "world model" with physical common sense and domain knowledge
FoxBrain 4.0A vision-language-action model able to control robots end to end

Expanded article table

Li said the institute expected to release the next-stage model, with vision, in 2026, as a step toward physical AI.[10] He gave production-line changeover as an example: a VLA model could reconfigure a robot arm quickly without a team of experts redesigning the process.[10] 劉尚青 of HHRI's AI Research Center said the main obstacle was action data: public robotics data is on the order of millions of tokens, and the gap to close is to the 15-trillion-token scale.[10]

Foxconn's COMPUTEX 2026 announcement (1 June 2026) said HHRI would introduce FoxBrain 2.0 at NVIDIA GTC Taipei. The company said the model "integrates NVIDIA Nemotron Omni architecture and multimodal training" and moves from "understanding" to "thinking, planning, and execution," enabling physical AI for its three platforms.[11] NVIDIA's GTC Taipei live blog, in a 1 June 2026 entry on sovereign AI, said Visionbay.ai "is building FoxBrain 2.0, a manufacturing foundation model for factory agents with a post-trained version of Nemotron 3 Nano Omni model on Visionbay.ai."[12] Nemotron 3 Nano Omni is the multimodal member of NVIDIA's Nemotron 3 family. The two companies attribute FoxBrain 2.0 differently: Foxconn credits HHRI, while NVIDIA names Visionbay.ai, the Foxconn cloud subsidiary on whose infrastructure the model is post-trained.[11][12] On 20 October 2026, DIGITIMES reported that Foxconn was expanding FoxBrain into vision-language models and vision-language-action models aimed at smart manufacturing, electric vehicles and robotics.[13]

Training infrastructure: the 2026 AI factory case study

In October 2026 NVIDIA published a customer case study, "Foxconn Industrializes Intelligence and Scales the Future With AI Factories." Foxconn posted a summary under the same title on 26 October 2026, and NVIDIA's AI infrastructure account on X promoted the study with the claim "1 trillion tokens. In just 12 days. That's how fast Foxconn trains its FoxBrain models."[14][15][16] The case study says Foxconn trains and refines its models on an NVIDIA DGX SuperPOD built from DGX GB200 systems (Blackwell GPUs connected by NVLink Switch in 72-GPU racks). It scales them out across the company on NVIDIA-Certified GB300 NVL72 systems and runs factory-floor digital twins and agents on NVIDIA RTX PRO servers.[14]

Reported resultDetail and basis
2.48x training speedupOn "120B parameter models" with DGX GB200, against Foxconn's previous H100-based systems. Foxconn-measured; NVIDIA notes it is consistent with its MLPerf Training v5.0 result of 2.5x Blackwell over Hopper on Llama 3.1 405B[14]
1 trillion tokens in 12 daysFoxconn's largest training jobs to date, from DGX SuperPOD job telemetry[14]
4,900 TFLOPS peak throughputPer Blackwell GPU, BF16/FP16 sparse Tensor Core throughput, as measured by Foxconn[14]
200+ facilitiesAI workloads orchestrated with NVIDIA Mission Control, Run:ai and NIM[14]
25% fewer scheduling hoursAutomated Production Scheduling (APS) cut labor hours for plant scheduling by 25%[14][15]
50% less engineering workloadSetpoint optimization for injection-molding process changeovers[14][15]
20+ defect root causesAddressed on surface-mount technology (SMT) lines, with a correction rate above 90%[14][15]

Expanded article table

The case study quotes Li, now titled founding director of the AI Research Center: "We achieved a 2.48x training speedup on our 120B parameter models compared to prior H100-based solutions."[14] It does not name the 120B model. Foxconn's own summary reports "2.48x faster training for 120-billion-parameter models" in its paragraph on the DGX SuperPOD training foundation, also without naming the model, and NVIDIA's post attributes the 12-day figure to FoxBrain training.[15][16] Kathy Yang, head of AI software business at Foxconn, described the DGX SuperPOD as the "Super Brain" where Foxconn trains and refines its most complex models.[14] The case study adds that Foxconn's AI factory is expanding with regional hubs in Kaohsiung and that Foxconn and NVIDIA are co-developing an "AI Token Ecosystem and Open Agent Framework" so other enterprises can build agents on Foxconn's platforms.[14] Taiwanese press coverage of the case study repeated the same figures and attributed them to NVIDIA.[17]

Timeline

DateEvent
10 March 2025HHRI announces FoxBrain (Llama 3.1 70B base, 120 H100 GPUs, about four weeks of training)[1][3]
20 March 2025Scheduled GTC 2025 session presenting FoxBrain[1]
13 May 2025Li outlines plans to open-source 8B and 70B models and a 1.x/2.0 roadmap[5]
18 May 2025NVIDIA describes the FoxBrain-powered Nurabot nursing robot[9]
19 May 2025Llama_3.1-FoxBrain-70B (v1.0 preview) repository created on Hugging Face[6]
31 July 2025FoxBrain v1.2 released on Hugging Face for academic and research use[7][8]
November 2025In Hon Hai Tech Day coverage, Li describes the four-stage roadmap to a VLA model[10]
1 June 2026Foxconn announces FoxBrain 2.0, based on NVIDIA Nemotron Omni, for GTC Taipei[11][12]
20 October 2026DIGITIMES reports FoxBrain's expansion into VLM and VLA models[13]
26 October 2026Foxconn and NVIDIA promote the DGX SuperPOD AI factory case study[14][15][16]

Expanded article table

See also

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27Hon Hai Research Institute, "Hon Hai Research Institute Launches Traditional Chinese LLM With Reasoning Capabilities," PR Newswire (via AAP), 10 March 2025. prod-ss.aap.com.au/...cision20250310ae37119
  2. ^Hon Hai Technology Group, "鴻海研究院推出第一版具有推理能力繁體中文大型語言模型" (Chinese-language press release), 10 March 2025. honhai.com/...1548
  3. ^1 ^2 ^3 ^4 ^5 ^6Wen-Yee Lee, "Foxconn unveils 'FoxBrain,' built on Nvidia GPUs to boost AI efforts," Reuters (via Yahoo Finance), 10 March 2025. finance.yahoo.com/...irst-large-language-072858118
  4. ^"Hon Hai launches its first advanced LLM 'FoxBrain'," Taipei Times (with CNA), 11 March 2025. taipeitimes.com/...2003833206
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7蘇文彬, "鴻海擬於5月底前開源繁中大語言模型FoxBrain," iThome, 13 May 2025. ithome.com.tw/...168922
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9FoxconnAI, "Llama_3.1-FoxBrain-70B" model card, Hugging Face (created 19 May 2025). huggingface.co/...Llama_3.1-FoxBrain-70B
  7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12FoxconnAI, "Llama_3.1-FoxBrain-70B-V1.2" model card and usage agreement, Hugging Face (created 31 July 2025). huggingface.co/...Llama_3.1-FoxBrain-70B-V1.2
  8. ^1 ^2 ^3 ^4吳季柔, "AI「首腦」FoxBrain！鴻海和輝達進軍美國製造靠它？," Global Views Monthly (遠見), 5 August 2025. gvm.com.tw/...123239
  9. ^1 ^2 ^3NVIDIA, "Foxconn Taps NVIDIA to Accelerate Physical and Digital Robotics for Global Healthcare Industry," NVIDIA Blog, 18 May 2025. blogs.nvidia.com/...foxconn-smart-hospital-robot
  10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9吳季柔 and 傅莞淇, "鴻海科技日》FoxBrain如何治矽谷AI巨頭的致命傷？," Global Views Monthly (遠見), 24 November 2025. gvm.com.tw/...126059
  11. ^1 ^2 ^3 ^4 ^5Hon Hai Technology Group, "Hon Hai Technology Group (Foxconn) Demonstrates AI-driven Platform Capabilities Across Robotics, Space, Healthcare At COMPUTEX 2026," 1 June 2026. foxconn.com/...2044
  12. ^1 ^2 ^3 ^4 ^5NVIDIA, "NVIDIA GTC Taipei at COMPUTEX: Live Updates on What's Next in AI," NVIDIA Blog (entry of 1 June 2026). blogs.nvidia.com/...-gtc-taipei-computex-2026-news
  13. ^1 ^2 ^3Ninelu Tu, "Foxconn's FoxBrain 2.0 sets sights on physical AI with VLM, VLA ambitions," DIGITIMES, 20 October 2026. digitimes.com/...foxconn-physical-ai-llm-data
  14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18NVIDIA, "Foxconn Industrializes Intelligence and Scales the Future With AI Factories" (customer case study), October 2026. nvidia.com/...foxconn-ai-factory-manufacturing
  15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Hon Hai Technology Group, "Foxconn Industrializes Intelligence and Scales the Future With AI Factories," 26 October 2026. foxconn.com/...2129
  16. ^1 ^2 ^3NVIDIA AI Infrastructure (@NVIDIAAIInfra), post on X, 26 October 2026. x.com/...2114813402228871201
  17. ^"工作量減 50%，鴻海合作 Nvidia 秀智造成果," TechNews (科技新報), 26 October 2026. finance.technews.tw/...nn-nvidia-smart-manufacture
  18. ^Zhi-Rui Tam, Ya-Ting Pai, Yen-Wei Lee et al., "An Improved Traditional Chinese Evaluation Suite for Foundation Model," arXiv:2403.01858, 2024. arxiv.org/...2403.01858
  19. ^Yen-Ting Lin and Yun-Nung Chen, "Taiwan LLM: Bridging the Linguistic Divide with a Culturally Aligned Language Model," arXiv:2311.17487, 2023. arxiv.org/...2311.17487
  20. ^TranNhiem, "FoxBrain_LLMs," GitHub repository. github.com/...FoxBrain_LLMs

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision · v1 · 3,137 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Model specifications, release history and the 2026 AI factory figures checked against Hon Hai Research Institute and Foxconn releases, the Hugging Face model cards, the NVIDIA case study, and Reuters, Taipei Times and Global Views coverage, October 2026.

Cite this page: AI Wiki. "FoxBrain." aiwiki.ai, updated 28 Oct 2026, fact-checked 28 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/foxbrain

Suggest edit

Where to go next

Topics:AI ModelsLarge Language ModelsReasoning Models

What links here

Foxconn (Hon Hai Precision Industry)