DeepSeek

High-Flyer, a Chinese hedge fund, spun off its AGI research lab as an independent company named DeepSeek, with founder Liang Wenfeng as CEO. The company focused on developing AI tools unrelated to High-Flyer's financial business.

DeepSeek released its first model, DeepSeek Coder, on November 2, 2023, followed by the DeepSeek-LLM series on November 29, 2023. Both were architecturally similar to Llama and released as open-weights.

DeepSeek released DeepSeek-Coder, a specialized code generation model trained on 2 trillion tokens of code data that achieved competitive performance against GPT-3.5 and CodeLlama on coding benchmarks. The model demonstrated DeepSeek's early focus on practical, high-performance models despite limited compute resources.

DeepSeek released V2, introducing multi-head latent attention (MLA) and Mixture-of-Experts architecture that significantly reduced KV cache size and memory usage, enabling more efficient inference.

DeepSeek released V3, claiming training cost of only $6 million — less than one-tenth of comparable models. The model used Mixture-of-Experts with 671B total parameters and demonstrated the ability to train competitive models under US chip export restrictions.

DeepSeek launched R1, a reasoning model comparable to OpenAI's o1, as a free app. By January 27, it surpassed ChatGPT as the most downloaded iOS app in the US, triggering an 18% drop in Nvidia's stock price — $600 billion in market value, the largest single-company decline in history.

DeepSeek released R1, a reasoning model that matched OpenAI's o1 performance on math and coding benchmarks while being trained at a fraction of the cost. R1's open-weight release and detailed training methodology triggered a global reassessment of AI efficiency, dubbed 'China's Sputnik moment' in the AI industry.

DeepSeek released V3.1 under the MIT License, featuring a hybrid architecture with thinking and non-thinking modes, surpassing prior models by over 40% on benchmarks like SWE-bench.

DeepSeek released DeepSeek-V3.2 and V3.2-Speciale, open-weight models the company said achieved world-leading reasoning results, including gold-medal-level performance on the IMO, CMO, ICPC and IOI 2025 competitions, alongside native tool-use 'thinking'. The release capped a run of sparse-attention innovations (DSA in September 2025) that cut long-context training and inference costs and kept DeepSeek the reference point for open-weight frontier efficiency ahead of the V4 generation.

DeepSeek released a preview of V4, including the 284-billion parameter V4-Flash and 1.6-trillion parameter V4-Pro, both featuring a one million token context window under the MIT License. V4 was adopted by Huawei and Cambricon for their chips.

On July 7, 2026, Reuters reported that DeepSeek was developing its own custom AI inference chip to reduce reliance on NVIDIA and Huawei hardware. The strategic move marked a significant vertical integration play by the Chinese AI company, which aimed to design its own silicon for inference workloads. The chip development represented DeepSeek's ambition to control its own hardware supply chain amid growing US-China technology export restrictions.

DeepSeek released V4 Flash 0731 as a public beta on the API, the same 284B/13B MoE architecture re-post-trained on agent data. Terminal-Bench 2.1 score jumped from 61.8% (preview) to 82.7%, surpassing V4-Pro-Preview (72.1%). Priced at $0.14/$0.28 per million tokens, it became the cheapest capable agentic model available. MIT-licensed weights published on HuggingFace, with native Responses API and Codex support.

DeepSeek resumed its second funding round seeking close to $8 billion at a valuation of approximately $74 billion, reportedly to fund a $50 billion data center in Inner Mongolia and a potential Shanghai IPO. The round follows a $7 billion raise at $52 billion valuation completed in late May 2026, making it one of the fastest consecutive funding rounds in AI history.

DeepSeek released V4 Pro 0813 as the general-availability flagship, a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters and a 1 million-token context window. The checkpoint delivers a 15.8-point improvement over the April preview on Terminal Bench 2.1 (87.9), placing it 0.1 points behind Anthropic Fable 5. DeepSeek simultaneously warned of significant API price increases, signaling the end of its aggressive undercutting strategy.

DeepSeek released V4.1-Flash, the smallest model in its new V4.1 architecture family, with vision support and aggressive pricing. DeepSeek reported that third-party tests put the compact model ahead of the larger V4-Pro on performance, cost, speed and total runtime, intensifying its pressure on U.S. frontier labs' pricing and prompting analysis of its disruption to the American AI market.

DeepSeek appointed GL Ventures partner Yan Wentao, a 1991-born investor known for early large-model and embodied-AI bets, as its first chief financial officer, and reportedly engaged CITIC Securities to prepare a potential IPO on Shanghai's STAR Market. The move ended a CFO vacancy that had persisted since early 2025 and marked the open-weights champion's shift from a research outfit toward a capitalized, listed company, as its financing scaled toward the 100-billion-yuan level.