Who Made DeepSeek V4? The Company and Team Behind the Model

Who Made DeepSeek V4? The Company and Team Behind the Model

DeepSeek V4 is the work of DeepSeek — formally Hangzhou DeepSeek Artificial Intelligence Co., Ltd. — a Chinese AI research company that has become one of the most influential players in global AI development. Here's a complete look at the organization, its history, and the philosophy that has driven the creation of one of the most capable open-weight AI models in the world.

DeepSeek: The Company

Full name: Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
Founded: 2023
Headquarters: Hangzhou, China (with research operations in multiple cities)
Parent company: High-Flyer Capital Management (a leading Chinese quantitative hedge fund)
Mission: Advancing AI research toward Artificial General Intelligence (AGI)

DeepSeek was founded as a subsidiary of High-Flyer Capital Management, one of China's most successful quantitative investment firms. High-Flyer's founders had the insight early on that building foundational AI capabilities was both strategically and commercially important — and that the best way to do so was to create a dedicated, well-funded AI research lab with an explicit research mission.

History and Key Milestones

DeepSeek's trajectory has been remarkably consistent: each major release has either matched or surpassed the leading models from OpenAI, Google, and Anthropic — at a fraction of the reported training cost.

The Team and Research Culture

DeepSeek is notable for its research-first culture. Unlike many AI labs that prioritize product launches and revenue growth, DeepSeek's team is known for:

• Publishing detailed technical reports alongside model releases
• Open-sourcing weights under permissive licenses (MIT)
• Sharing architectural innovations that other labs can build upon
• Explicitly stating a long-term commitment to AGI research over short-term commercial optimization

The DeepSeek V4 technical report (available on HuggingFace) is a thorough academic document detailing the Hybrid Attention Architecture, mHC, Muon optimizer, and training pipeline — not a marketing brochure.

The team maintains official communication channels via:

• Twitter/X: @deepseek_ai
• GitHub: github.com/deepseek-ai
• HuggingFace: huggingface.co/deepseek-ai
• Contact: service@deepseek.com

DeepSeek's Philosophy: Longtermism and AGI

In its official communications, DeepSeek explicitly embraces a philosophy of longtermism — prioritizing research and development choices that maximize long-term value over short-term metrics.

From the V4 announcement: "We remain committed to longtermism, advancing steadily toward our ultimate goal of AGI."

This manifests in several specific ways:

1. Open-sourcing everything: Rather than monetizing model weights, DeepSeek gives them away under MIT — betting that ecosystem growth and talent attraction are more valuable long-term
2. Research transparency: Publishing detailed technical reports helps the field advance, even if it gives competitors insight into DeepSeek's innovations
3. Efficiency focus: DeepSeek consistently pushes the frontier of compute efficiency, believing that making powerful AI cheaper is the highest-leverage contribution they can make

Training Infrastructure: Hardware Independence

One particularly notable aspect of DeepSeek V4 is its hardware context. DeepSeek trained V4 on Huawei Ascend 950PR GPUs — Chinese-manufactured AI accelerators — rather than NVIDIA A100s or H100s. This reflects both practical necessity (US export restrictions limit Chinese companies' access to NVIDIA's most advanced chips) and strategic intent (developing domestic AI capabilities).

The fact that V4-Pro — trained on Ascend hardware — achieves Codeforces rating 3206 and leads on LiveCodeBench demonstrates that world-class AI training is now achievable without NVIDIA's most advanced chips.

For local inference, however, DeepSeek V4 runs well on standard NVIDIA hardware — the H100 ecosystem for professional deployment and NVIDIA consumer GPUs (with quantization) for smaller-scale use.

DeepSeek's Relationship with the Broader AI Ecosystem

DeepSeek collaborates with and influences the global AI ecosystem in several ways:

• Integration with western tools: V4 is natively compatible with Claude Code, OpenClaw, and OpenCode
• HuggingFace partnership: All models are distributed through HuggingFace, the de facto standard for open-weight model distribution
• API compatibility: V4's API matches both OpenAI and Anthropic formats, making it a drop-in replacement for western models
• Community engagement: DeepSeek actively engages with open-source communities through GitHub issues, HuggingFace discussions, and social media

Notable Facts About DeepSeek

• DeepSeek has consistently achieved results that surprised the western AI industry — with R1 in January 2025 being perhaps the most significant shock, demonstrating frontier reasoning capabilities at dramatically lower reported training costs
• DeepSeek operates with remarkable transparency for an AI company — publishing technical reports, sharing training details, and communicating directly with researchers
• All official news and updates come through their official accounts — DeepSeek explicitly warns users to rely only on official channels for DeepSeek news

Connecting With DeepSeek V4

Whether you're building AI applications on Framia.pro or developing your own products, DeepSeek V4 is accessible today through:

• API: api.deepseek.com (using your DeepSeek API key)
• Web chat: chat.deepseek.com
• Open weights: huggingface.co/deepseek-ai
• Documentation: api-docs.deepseek.com

Conclusion

DeepSeek is one of the most consequential AI labs in the world — not despite being Chinese, but as a direct product of its research culture, long-term thinking, and commitment to open science. DeepSeek V4 is the clearest expression yet of what happens when a team with serious compute resources, research depth, and a genuine commitment to advancing AI for humanity's benefit builds without short-term commercial constraints.