High-Flyer AI

What High-Flyer AI does

High-Flyer AI (幻方) is a privately-held Chinese AI research organization founded in 2015 and headquartered in Hangzhou, operating as a dual-entity holding structure comprising a quantitative hedge fund business and the DeepSeek AI research subsidiary. The quantitative fund business, operated through Zhejiang Jiu Zhang Asset Management, Ningbo High-Flyer Quant Investment, and High-Flyer Capital Management (Hong Kong), serves 10,000+ high net worth clients and institutional investors using AI-driven trading strategies accumulated since 2008, with consistent Golden Bull Award recognition and top-tier AUM in China's asset management industry. The company has accumulated over 10PB of proprietary market, financial, and macroeconomic data across dozens of daily-updated sources.

The DeepSeek subsidiary, established in 2023 when High-Flyer pivoted from pure AI-driven trading toward Artificial General Intelligence research, develops open-source large language models including the V3, V3.1, R1, and V4 series using a 685 billion parameter mixture-of-experts architecture. The company's proprietary technical stack centers on the Fire Flyer (萤火) GPU training cluster family: Fire Flyer 1 (approximately 500 GPUs, nearly 200 million yuan invested in 2019) and Fire Flyer 2 (1 billion yuan invested in 2021, achieving 96% cluster utilization, 85% GPU utilization, and 8 TB/s read bandwidth via the 3FS file system). Surrounding infrastructure includes HAI-LLM (four-way parallelism training tool), hfreduce (CPU-based gradient aggregation), haiscale (parallel training toolkit), FFRecord (training data format), haiprof (profiling tool), hfai.nn (custom CUDA operators with 6-10x LSTM speedups), and CC_Cleaner (Common Crawl processing pipeline). The R1 model was trained for approximately $6 million on 2,000 Nvidia H800 GPUs versus hundreds of millions spent on comparable Western frontier models.

Revenue mechanics operate across three streams: subscription-based quantitative fund management fees, usage-based API access for DeepSeek models at significantly undercut pricing, and open-source distribution prioritized over near-term commercial monetization. Distribution channels include direct enterprise sales for the quant business, platform integrations with Microsoft (Copilot Cowork), AWS, and Nvidia for DeepSeek, and global open-source downloads via GitHub and Hugging Face. In 2026, DeepSeek completed its first external fundraising round at a $50 billion valuation led by the state-linked National AI Investment Fund with Tencent participation, though US government and military entities are explicitly excluded from the addressable market.

Name

High-Flyer AI

Legal name

幻方量化 (High-Flyer Quant)

Website

https://high-flyer.cn

Company type

Private

Founded year

2015

Operating status

Operating

Headcount range

101–250 employees

Short description

High-Flyer AI (幻方) is a Hangzhou-based dual-entity organization operating one of China's largest AI-powered quantitative hedge funds serving 10,000+ high net worth clients, alongside the DeepSeek AI research subsidiary developing open-source frontier large language models at a fraction of Western training costs.

Ownership category

akta.pro rank

Product category

Quantitative Investment & AI Infrastructure

NAICS

Portfolio Management and Investment Advice (52394), Computer Systems Design and Related Services (54151)

SIC

Services-Computer Integrated Systems Design (7373), Investment Advice (6282)

akta.pro primary industry

Model Hosting, Serving & Inference Platforms (HDAAACAB)

akta.pro secondary industries

End-to-End Enterprise AI Platforms (MLOps & Model Lifecycle Management) (HDAEANAA), AI Compiler, Runtime & Kernel Optimization Software (CUDA/ROCm/XLA, graph compilers) (HDAAAAAI), Enterprise Foundation Model Integration & APIs (Connectors, Governance, Deployment) (HDAAACAO), Model Compression & Optimization (Quantization, Distillation, Pruning) (HDAAACAL), AI/ML Engineering & Model Development Services (BPAEAHAF)

Keywords

• Quantitative investment
• AI training platform
• Large language models
• Deep learning infrastructure
• Hedge fund management

Headquarters

HQ city

Hangzhou

HQ country

China

HQ region

Asia

Offices3 records

Markets served

GTM type

B2B and B2C

Offering type

Software

Cost components

Technology or R&D, Personnel, Infrastructure, Operations

Revenue model

1. Quantitative Fund Management: Asset management services providing AI-driven quantitative trading to high net worth clients and institutional investors. Company manages assets through Zhejiang Jiuzhang Asset Management and Ningbo High-Flyer Quant Investment, consistently winning Golden Bull Awards.
2. AI Model API Access: DeepSeek provides API access to its AI models at significantly lower pricing than Western competitors. Models like V4 Pro and V4 Flash reportedly rival Western competitors at lower costs.
3. Open Source Community: DeepSeek prioritizes open-source AGI development over commercial revenue generation, releasing models like V4 Pro and V4 Flash to build developer ecosystem and market influence.

Pricing tiers

Go-to-market motion3 records

Distribution channels3 records

Marketing channels4 records

Core offering

High-Flyer AI operates a dual business combining AI-powered quantitative investment fund management (serving 10,000+ high net worth clients and institutional investors) with proprietary AI infrastructure research and development. The company builds and open-sources deep learning training tools (HAI-LLM, HAI Platform, 3FS, hfreduce) and develops large language models through its subsidiary DeepSeek (V3, R1).

Product overview

High-Flyer AI (幻方量化) operates as a dual business entity combining AI-powered quantitative investment fund management with AI infrastructure research and development. The company's product portfolio centers on its self-developed AI training ecosystem comprising the 萤火二号 (Fire Flyer Two) deep learning training platform with task-level time-sharing architecture, supplemented by open-source tools including HAI-LLM for large model training, HAI Platform for GPU cluster resource management, haiscale for parallel training, hfreduce for multi-card communication, 3FS high-speed file system, FFRecord data format, haiprof performance profiling, hfai.nn optimized operators, and CC_Cleaner data processing. DeepSeek, the company's AI model subsidiary, develops open-source foundation models including V3, R1, V3.1, and V4 Pro/Flash that compete with Western frontier AI systems. The quantitative fund business uses AI technology for investment management serving 10,000+ high-net-worth clients.

Differentiator

Problem solved

Functional benefit

Brands

• 萤火一号 (Fire Flyer 1): First generation AI training cluster with approximately 500 GPUs, invested nearly 200 million yuan
• 萤火二号 (Fire Flyer 2)
• 3FS
• HAI Platform
• HAI-LLM
• hfreduce
• FFRecord

Products and services

• Quantitative Fund (量化基金) AI-powered quantitative investment fund management using neural networks for market data processing, natural language understanding, and financial behavior analysis. Serves 10,000+ high-net-worth clients and institutional investors with multi-year Golden Bull Award recognition.
• Fire Flyer Two (萤火二号) Second-generation deep learning training cluster with task-level time-sharing architecture, 800-port switch interconnection, and core expansion subtree architecture. Achieves 96% cluster usage rate and 85% GPU utilization with 8.0 TB/s read and 500 GB/s write bandwidth.
• HAI-LLM Efficient and lightweight large model training tool implementing four parallel training methods (ZeRO data parallelism, pipeline parallelism, tensor model parallelism, and sequence parallelism). Supports training of trillion-parameter models across thousands of GPUs.
• HAI Platform Open-source AI training platform for GPU cluster resource management featuring task-level time-sharing scheduling. Supports 1500+ compute nodes with 95%+ daily compute utilization and 75%+ GPU utilization.
• 3FS (Fire-Flyer File System) Self-developed high-speed distributed file system optimized for AI training sample data reading, using Direct IO and RDMA Read for ultra-high read bandwidth up to 7TB/s throughput.
• haiscale Lightweight high-performance parallel training toolkit integrating DDP, FSDP, and pipeline parallelism algorithms with hfreduce communication backend for better multi-GPU scaling performance than PyTorch native tools.
• hfreduce Self-developed high-performance multi-card parallel communication tool using CPU for gradient aggregation instead of NCCL, reducing PCIe traffic by half and avoiding GPU gradient computation overhead.
• FFRecord High-performance binary data format for training samples that supports random batch reading, integrates with PyTorch Dataset and DataLoader interfaces, and is optimized for 3FS storage performance.
• DeepSeek R1 Open-source large language model demonstrating performance comparable to Western AI systems (OpenAI, Anthropic) at approximately $6 million training cost, using 2,000 Nvidia H800 GPUs and mixture-of-experts architecture.
• DeepSeek V3.1 Open-source large language model with 685 billion parameters using mixture-of-experts architecture, achieving performance rivaling OpenAI and Anthropic systems at a fraction of the cost.

Quantifiable outcome

• DeepSeek R1 trained for ~$6M vs $80-100M required for comparable proprietary models
• +5 more outcomes

Named customers1 record

Segments3 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability6 records

Feature8 records

Partnerships

Three partnerships are on record, tiered major.

• MicrosoftmajorTechnology or IntegrationMicrosoft is reportedly considering offering DeepSeek's V4 model through its Copilot Cowork platform, representing significant validation of DeepSeek's technical capabilities by a leading global AI provider.
• AWS (Amazon Web Services)majorTechnology or IntegrationAWS has secured platform integration with DeepSeek models, enabling cloud customers to access DeepSeek's cost-efficient AI capabilities through AWS infrastructure.
• NvidiamajorTechnology or IntegrationNvidia has integrated DeepSeek models into its platform ecosystem, providing access to optimized AI capabilities for Nvidia GPU users and validating DeepSeek's performance on Nvidia hardware.

Scale indicators11 records

Recent moves6 records

Expansion highlights6 records

Direct peers

• Cohere: Enterprise-focused LLM provider with API and private deployment offerings. Comparable to DeepSeek on enterprise go-to-market and serving similar developer audiences, though operating as a standalone company without the quant fund financial backing and Chinese cost advantages.
• Anthropic: Frontier LLM developer (Claude) with safety-focused positioning and enterprise API revenue. Directly comparable to DeepSeek on model capability targets and enterprise distribution, though operating from a different geopolitical footprint and with substantially higher training costs.
• Baidu (ERNIE): Chinese tech giant's foundation model division (ERNIE/Wenxin). Competes with DeepSeek for Chinese enterprise and government customers, with broader application integration across Baidu's product ecosystem but historically closed-source distribution.
• Mistral AI: European open-weight LLM developer with strong enterprise focus. Highly comparable to DeepSeek on open-source distribution strategy, API pricing, and serving European/global enterprise customers, though at a smaller scale and without the quant fund parent backing.
• Alibaba (Qwen): Chinese tech giant's foundation model division releasing Qwen open-source LLMs. Direct competitor in the Chinese and global open-source LLM market, with similar cost-efficient training approach and enterprise cloud distribution through Alibaba Cloud.
• Meta AI (LLaMA): Meta's AI research division releasing LLaMA open-weight models. Comparable open-source foundation model strategy with global enterprise adoption, though operating from a hyperscaler parent with substantially more compute resources than DeepSeek.
• OpenAI: Frontier foundation model developer (GPT-4/o-series) serving enterprise APIs and consumer products. Direct competitor to DeepSeek on capability, with comparable revenue model (API + enterprise) but vastly larger training budgets and Western enterprise/government market access.

Broad incumbents

• Google DeepMind: Alphabet's AI research arm developing Gemini models. Operates broadly in the foundation model space with frontier capability targets, but as part of a much larger portfolio (cloud, search, consumer) that does not specialize in the cost-disruptor positioning that defines DeepSeek.
• Citadel Securities: Global quantitative trading firm and market maker. Highly comparable to High-Flyer's quant fund business on AI-driven trading strategies, institutional client base, and asset management scale, though not involved in foundation model development.

Emerging players

• xAI (Grok): Elon Musk's AI company developing Grok foundation models with consumer and enterprise distribution. Comparable as a frontier model competitor with significant compute resources, but with strategic priorities and distribution model heavily tied to the X platform rather than open-source enterprise adoption.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks6 records

Key highlights7 records

Customer concentration

• GitHubgithub.com/HFAiLab
• Xx.com/HighFlyer_Quant
• Facebookfacebook.com/HighFlyerAI

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Subsidiaries4 records

Funding overview

Funding rounds

Investors

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about High-Flyer AI