DeepSeek

DeepSeek is a Chinese artificial intelligence research company founded in 2023 by Liang Wenfeng and headquartered in Hangzhou. The company develops large language models distributed under open licenses, including the DeepSeek-V3, DeepSeek-R1, and DeepSeek-V4 families, and is widely regarded as the leading open-weights frontier developer based in China. DeepSeek's January 2025 release of R1 disrupted prevailing assumptions about the cost of frontier capability, and its April 2026 V4 release pushed the disruption further by combining a 1.6-trillion-parameter mixture-of-experts architecture with full deployment on Chinese-built Huawei Ascend silicon.

At a glance

• Founded: April 2023 in Hangzhou, China, as a spinout of the High-Flyer quantitative hedge fund.
• Status: Private. Self-funded through April 2026; first external fundraising round reported as in process.
• Funding: Internally funded through High-Flyer profits since founding. Reported $300 million round in process at a $10 billion-plus valuation as of April 2026, the company's first external raise.
• CEO: Liang Wenfeng (founder; co-founder of High-Flyer)
• Other notable leadership: Engineering and research leadership team has not been publicly profiled in detail; the company maintains an unusually low public-relations footprint relative to peer frontier labs.
• Open weights: Yes. Every published DeepSeek model has been released under an open license, with weights and inference code on Hugging Face and GitHub.
• Flagship models: DeepSeek-V4 Pro and DeepSeek-V4 Flash (April 2026 preview release). Prior flagships include DeepSeek-V3 (December 2024) and DeepSeek-R1 (January 2025).

Origins

DeepSeek was established in April 2023 as a research body funded by High-Flyer, a Chinese quantitative hedge fund that Liang Wenfeng co-founded in February 2016 with classmates from Zhejiang University. High-Flyer's assets under management surpassed RMB 100 billion in 2021, and the firm built large GPU clusters in 2020 and 2021 for trading research that provided the compute base from which DeepSeek's first models were trained.

High-Flyer announced DeepSeek as an artificial general intelligence research body explicitly outside the trading business. Liang has stated in Chinese-language interviews that DeepSeek was structured to run on the hedge fund's compute and capital so that researchers could focus on architectural and training research without the fundraising and product-launch pressures that constrain peer labs.

The first DeepSeek models in 2023 and early 2024 were dense and mixture-of-experts language models in the 7-billion to 67-billion-parameter range, released open-weight and benchmarked against contemporary open releases (Meta AI / FAIR's Llama family, Mistral AI's open releases, Alibaba's Qwen line). The company also shipped specialized DeepSeek-Coder and DeepSeek-Math models that became reference open-weights options for code and mathematical reasoning.

The late-2024 and early-2025 releases changed the company's profile globally. DeepSeek-V3 launched in December 2024 with 671 billion total parameters in a mixture-of-experts configuration (37 billion active) and reported training cost in the range of $5 million to $6 million using H800 GPUs. DeepSeek-R1 followed on January 20, 2025, applying large-scale reinforcement learning to V3 to produce a reasoning model that matched OpenAI's o1 on math, code, and reasoning benchmarks. R1 became the most-downloaded free app on the US Apple App Store within weeks of release, an event covered as the "DeepSeek shock" because of the corresponding sell-off in US AI hardware stocks and the reframing of the cost-of-frontier-AI narrative.

DeepSeek-V4, previewed on April 24, 2026, is the company's first frontier-tier release built and shipped in close integration with Huawei's Ascend AI chips rather than the Nvidia-CUDA ecosystem that has dominated frontier training to date. The V4 release coincided with the company's first reported external fundraising effort.

Mission and strategy

DeepSeek's stated mission, as expressed in Liang Wenfeng's public Chinese-language interviews and in the company's research papers, is to advance fundamental research toward artificial general intelligence with open-source releases as the primary distribution channel. The company operates with a small team relative to its capability profile and has not adopted the consumer-product or enterprise-API positioning that characterizes US frontier labs.

The strategy combines three threads. First, architectural research focused on training and inference efficiency at frontier scale, including innovations in mixture-of-experts routing, multi-head latent attention, multi-token prediction, and mixed-precision training documented in detailed technical reports with each release. Second, open-weights release of every model under permissive licenses through Hugging Face and GitHub. Third, integration with Chinese cloud and chip infrastructure, accelerating in 2026 with V4's Huawei Ascend deployment.

The competitive premise is that algorithmic and engineering innovation can compress the cost of frontier capability faster than the brute-force scaling pursued by US frontier labs, and that open-weights distribution can produce capability proliferation that closed-weights vendors cannot match on cost or developer ergonomics. The premise has been validated by R1's reception in early 2025 and by V4's benchmark profile in April 2026. The Huawei Ascend integration in V4 extends the strategy beyond software efficiency: by demonstrating frontier-tier training without Nvidia CUDA, DeepSeek positions itself as a structural participant in the Chinese AI hardware ecosystem.

Models and products

• DeepSeek-V4 Pro and V4 Flash. Mixture-of-experts language models released in preview April 2026. V4 Pro reports 1.6 trillion total parameters with 49 billion active and a 1-million-token context window. V4 Flash is the smaller and lower-cost sibling. Both shipped on Huawei Ascend infrastructure.
• DeepSeek-R1 and R1-Zero. Released January 2025. Reasoning-optimized model trained with reinforcement learning on V3. R1 became one of the most-cited open-weights models of 2025 and forms the lineage for community fine-tunes and distillations.
• DeepSeek-V3 and V3.2. Released December 2024 and 2025. The 671-billion-parameter mixture-of-experts foundation model that anchored the late-2024 and 2025 release cycle.
• DeepSeek-Coder, DeepSeek-Math, DeepSeek-VL. Specialized open-weights lines for code generation, mathematical reasoning, and vision-language understanding.

DeepSeek distributes weights through Hugging Face and GitHub. The company also operates a hosted chat product and a developer API at prices substantially below US frontier-lab equivalents.

Benchmarks and standing

As of April 2026, DeepSeek-V4 Pro is reported in the top-ten range across the major standardized benchmarks. The Artificial Analysis Intelligence Index places V4 Pro at rank 8 with a composite score of 51.51. On LMArena's general ELO leaderboard, V4 Pro is ranked 5; on the coding ELO leaderboard, rank 3 with an ELO of 1287; on the vision ELO leaderboard, rank 7. On SWE-bench Verified, V4 Pro reports a score of 64.2 (rank 3). On GPQA Diamond, 82.1 (rank 6). On HumanEval+ Pass Rate, 91.2 (rank 3). On the ARC-AGI Challenge, 79.5 (rank 5). On AIME 2025, 85.0 (rank 5).

These benchmark positions are point-in-time, taken at the V4 preview release. Benchmark leadership rotates on the timescale of weeks given the release cadence of competing labs through 2026, but DeepSeek's V4 places it in the leading group of open-weights frontier models alongside Meta AI / FAIR's Llama family and ahead of most peer Chinese open-weights releases.

The benchmark profile underwrites the strategic claim that the company's training-efficiency research has compressed the gap between open-weights and closed-weights frontier capability. Whether that gap continues to narrow in subsequent releases is one of the most-watched questions in 2026 industry coverage.

Leadership

As of April 2026, DeepSeek's senior leadership includes:

• Liang Wenfeng, founder and Chief Executive Officer. Co-founder of High-Flyer, the quantitative hedge fund that funds DeepSeek. Public face for the company's research and strategic direction. Has given several Chinese-language interviews on DeepSeek's research thesis but maintains a low international media profile relative to peer frontier-lab CEOs.

The company's research and engineering leadership team has not been publicly profiled in the depth that is typical of US frontier labs. Researchers credited on DeepSeek papers include senior contributors recruited from Chinese university programs (Tsinghua, Peking, Zhejiang) and from Chinese technology companies. The compact team size (reported under 200 people in 2025) is unusual for a frontier-capable lab and is a frequently cited element of the cost-efficiency story.

Notable external supporters mentioned in industry coverage include Qihoo 360 founder Zhou Hongyi, who has commented publicly in support of the company's strategy, though there is no public record of a formal investor or operating role.

Funding and backers

DeepSeek's funding model through April 2026 has been unusual among frontier labs. The company has been funded entirely by High-Flyer profits since founding, with no external venture capital, strategic-corporate, or sovereign-fund participation. Liang Wenfeng holds a reported 84.3 percent indirect ownership of DeepSeek through High-Flyer-affiliated investment vehicles plus a 1 percent direct holding, giving him operational and economic control.

In April 2026, DeepSeek began discussions with external investors for a reported $300 million funding round at a $10-billion-plus valuation. This would be the company's first external raise and reflects the cost pressures of training V4-scale models even with the company's documented efficiency advantages. The round had not closed and the lead investors had not been publicly disclosed as of April 2026. The transition to external fundraising is consistent with the scale of compute investment required for V4 and successor model training.

Industry position

DeepSeek occupies a distinctive position in the global AI landscape. The combination of frontier-tier benchmark capability, open-weights distribution, and Chinese national-strategic positioning produces a profile no US lab matches and no other Chinese lab matches at the same capability tier. Industry coverage frequently characterizes DeepSeek as the most strategically significant Chinese AI lab of the 2024 to 2026 period. The January 2025 R1 release was characterized by US press coverage as a national-competitiveness event, with associated policy debate in Washington over export controls and the structure of US AI capability investment.

Strategic risks include continued US export-control tightening that could constrain access to advanced Nvidia chips through any remaining channel, intensifying domestic Chinese competition from Alibaba, Tencent, ByteDance, and Moonshot as those labs increase capability investment, and the operational complexity of transitioning training and inference infrastructure to Huawei Ascend at scale. Strategic strengths are the capability tier, the open-weights distribution moat, the cost-efficiency demonstrated across releases, and the structural position as the principal Chinese open-weights frontier developer.

Competitive landscape

DeepSeek competes with several frontier and Insurgent labs:

• OpenAI. Direct frontier-capability competitor on closed-weights side. OpenAI's flagship and reasoning lines (GPT-5 family, o-series) are the principal benchmark targets that DeepSeek's V4 and R1 successors are positioned against.
• Anthropic. Frontier competitor on capability and on enterprise-AI deployment. Anthropic's coding capability through Claude Code is the principal closed-weights coding leader against DeepSeek's open-weights coding releases.
• Google DeepMind. Frontier competitor through the Gemini family. Less direct competitive overlap given DeepMind's distribution through Google products rather than as a separate API.
• Meta AI / FAIR. Direct open-weights competitor through the Llama family. Llama and DeepSeek are the two dominant open-weights frontier lines in 2025 and 2026; DeepSeek's V4 Pro and Llama's most recent releases compete head-to-head for the open-weights leaderboard position.
• Mistral AI. Open-weights competitor based in Europe; broadly smaller in scale than DeepSeek's flagship releases but competing for the same open-weights developer ecosystem.
• Reflection AI. Has explicitly positioned itself as "America's open frontier AI lab" with DeepSeek named as the principal Chinese open-weights frontier competitor. Reflection's planned 2026 open-weights frontier model is positioned to challenge DeepSeek-tier capability from a US base.
• Alibaba Qwen, Moonshot AI, Z.AI / Zhipu, MiniMax, ByteDance Seed, Tencent Hunyuan. Peer Chinese frontier labs. Each pursues a distinct strategy combination (cloud distribution for Alibaba, agent-product focus for Moonshot, enterprise positioning for Z.AI, multimodal scaling for MiniMax, internal-product distribution for ByteDance). DeepSeek competes with this cohort for engineering talent and for the open-weights leadership position within China.

Outlook

Several open questions affect DeepSeek's trajectory in 2026 and 2027:

• The full release of DeepSeek-V4 Pro and Flash beyond preview, including final benchmark profile, throughput on Huawei Ascend, and production-API pricing.
• The closure of the reported $300 million external funding round and the identity of lead investors. Sovereign-fund or strategic-corporate participation would be a structurally significant signal.
• The trajectory of Huawei Ascend integration. If V4 production deployment runs at scale on Ascend with comparable performance to Nvidia-based deployment, the strategic significance extends beyond DeepSeek to the Chinese AI hardware ecosystem.
• The competitive response from US frontier labs and from open-weights peers (Llama, Reflection AI's planned 2026 release).
• US policy responses on export controls and on intellectual-property allegations, which may shape the export-control environment for successor releases.
• The pace of successor model releases through 2026 and 2027.

Sources

• TechCrunch: DeepSeek previews new AI model that 'closes the gap' with frontier models. April 2026 V4 preview coverage.
• Fortune: DeepSeek unveils V4 model, with rock-bottom prices and close integration with Huawei's chips. V4 release and Huawei Ascend integration.
• CNBC: China's DeepSeek releases preview of long-awaited V4 model. V4 preview release coverage.
• MIT Technology Review: Three reasons why DeepSeek's new model matters. Strategic context for V4.
• Tech Startups: DeepSeek seeks $300M in first fundraise at $10B+ valuation as AI costs surge. April 2026 funding round context.
• DeepSeek-R1 Release announcement. Official R1 release notes from January 2025.
• Wikipedia: DeepSeek. Comprehensive company history reference.
• Wikipedia: High-Flyer. Parent hedge-fund history.
• Wikipedia: Liang Wenfeng. Founder biographical reference.