V4 Model Updates with Agentic Improvements

• DeepSeek-V4-Flash-0731 (GA release) — Released July 31, 2026, this is the official production-ready version of the Mixture-of-Experts model (284B total parameters, 13B active, 1M token context). It was re-post-trained specifically for agentic coding tasks and now exceeds the V4-Pro-Preview on every agentic benchmark DeepSeek publishes .

• V4-Pro Preview enhancements — The V4 Pro now preserves reasoning history across tool-result rounds and new user messages (previously discarded), and Nvidia's docs list it as suited for "agentic AI applications, tool use scenarios, and complex problem-solving" .

• API updates — The V4-Flash API now supports native Responses API and Codex-focused compatibility . DeepSeek's API also added Anthropic-format compatibility alongside its existing OpenAI-compatible endpoints .

Company Size, Funding, and Commercial Transformation

• Workforce — DeepSeek had roughly 160 employees as of mid-2026, but is pursuing an aggressive plan to double headcount across all departments .

• Maiden funding round (~$7.4B) — In June 2026, DeepSeek closed its first-ever external fundraising round, raising approximately $7.4 billion (50 billion yuan) at a valuation between $50–59 billion, led by Tencent and CATL as the largest external investors. Founder Liang Wenfeng maintained control via non-voting stakes .

• Second round (~$8B, resumed August 6) — DeepSeek paused a planned follow-on round in late July after founder remarks leaked, then resumed on August 6 seeking close to $8 billion, with Monolith Management in talks to participate .

• Valuation trajectory — From ~$50B (June) to a planned ~$74B round (July), with an onshore IPO targeted for 2027 on Shanghai's STAR Market .

U.S. Export Controls and Challenges

DeepSeek has operated under U.S. chip export restrictions that limit Chinese AI companies' access to advanced NVIDIA GPUs. The company's efficiency innovations — including its Mixture-of-Experts architecture and post-training breakthroughs — were partly born from necessity due to these constraints. The V4-Flash's ability to outperform larger models on agentic benchmarks while using fewer active parameters is seen as a direct response to hardware limitations .

Broader Context: Chinese AI Competitive Rise

DeepSeek's trajectory mirrors a broader wave of Chinese AI labs (including Alibaba's Qwen, Baidu's Ernie, and Zhipu AI) that have closed the gap with U.S. frontier models. DeepSeek's open-source strategy and dramatically lower API pricing have pressured U.S. competitors on cost, while its agentic coding push directly targets the most commercially valuable AI application layer — code generation — where Anthropic's Claude Code and GitHub Copilot currently lead .

V4 Model Updates with Agentic Improvements

• DeepSeek-V4-Flash-0731 (GA release) — Released July 31, 2026, this is the official production-ready version of the Mixture-of-Experts model (284B total parameters, 13B active, 1M token context). It was re-post-trained specifically for agentic coding tasks and now exceeds the V4-Pro-Preview on every agentic benchmark DeepSeek publishes .

• V4-Pro Preview enhancements — The V4 Pro now preserves reasoning history across tool-result rounds and new user messages (previously discarded), and Nvidia's docs list it as suited for "agentic AI applications, tool use scenarios, and complex problem-solving" .

• API updates — The V4-Flash API now supports native Responses API and Codex-focused compatibility . DeepSeek's API also added Anthropic-format compatibility alongside its existing OpenAI-compatible endpoints .

Company Size, Funding, and Commercial Transformation

• Workforce — DeepSeek had roughly 160 employees as of mid-2026, but is pursuing an aggressive plan to double headcount across all departments .

• Maiden funding round (~$7.4B) — In June 2026, DeepSeek closed its first-ever external fundraising round, raising approximately $7.4 billion (50 billion yuan) at a valuation between $50–59 billion, led by Tencent and CATL as the largest external investors. Founder Liang Wenfeng maintained control via non-voting stakes .

• Second round (~$8B, resumed August 6) — DeepSeek paused a planned follow-on round in late July after founder remarks leaked, then resumed on August 6 seeking close to $8 billion, with Monolith Management in talks to participate .

• Valuation trajectory — From ~$50B (June) to a planned ~$74B round (July), with an onshore IPO targeted for 2027 on Shanghai's STAR Market .

U.S. Export Controls and Challenges

DeepSeek has operated under U.S. chip export restrictions that limit Chinese AI companies' access to advanced NVIDIA GPUs. The company's efficiency innovations — including its Mixture-of-Experts architecture and post-training breakthroughs — were partly born from necessity due to these constraints. The V4-Flash's ability to outperform larger models on agentic benchmarks while using fewer active parameters is seen as a direct response to hardware limitations .

Broader Context: Chinese AI Competitive Rise

DeepSeek's trajectory mirrors a broader wave of Chinese AI labs (including Alibaba's Qwen, Baidu's Ernie, and Zhipu AI) that have closed the gap with U.S. frontier models. DeepSeek's open-source strategy and dramatically lower API pricing have pressured U.S. competitors on cost, while its agentic coding push directly targets the most commercially valuable AI application layer — code generation — where Anthropic's Claude Code and GitHub Copilot currently lead .