DeepSeek

DeepSeek is a Chinese AI company founded in July 2023 in Hangzhou by Liang Wenfeng, who previously co-founded High-Flyer Capital, a quantitative hedge fund. The company built on High-Flyer's existing computing infrastructure and quantitative-finance talent to pivot into foundational AI research. DeepSeek's stated mission centers on realizing artificial general intelligence, with an emphasis on foundational research rather than immediate commercialization, and the company has committed to open-sourcing its models to encourage transparency and community collaboration.

Notably, the company operated with remarkable capital efficiency, running on only 150–200 employees and relying on no external venture capital until 2026.

#The "DeepSeek Shock" of January 2025

DeepSeek became globally famous almost overnight. It rose to international prominence in January 2025 after releasing its mobile chatbot app alongside the large language model DeepSeek-R1. Released on January 10, the app became the most-downloaded app on Apple's U.S. App Store within weeks and ranked among the top downloads on Google Play. Its efficiency, affordability, and transparency relative to American AI competitors triggered a sharp selloff in U.S. tech stocks.

#Technical Approach

DeepSeek's models are built around a few core architectural ideas designed to cut training and inference costs dramatically:

• Mixture-of-Experts (MoE) architecture: In models like DeepSeek-V3 (671 billion total parameters), only around 37 billion parameters activate for any given task, keeping inference fast and cheap.
• Sparse attention: DeepSeek Sparse Attention reduces inference costs by roughly 70% for long-context inputs.
• Visible chain-of-thought reasoning: The models expose their step-by-step reasoning process, letting users inspect how an answer was derived.
• Open licensing: Model weights have been released under an MIT license, paired with a fully free chatbot, which significantly lowered the barrier to adoption for millions of users worldwide.

#Model Lineage

Based on independent tracking of DeepSeek's releases:

• V1 — dense models (7B/67B), 4K context
• V2 — MoE + latent attention, 236B total/~21B active, 128K context (with a lighter "V2-Lite" variant)
• V3 — 671B total/~37B active MoE, 256 experts, trained on roughly 14 trillion tokens
• R1 (Jan 2025) — reasoning-focused model; it scores 79.8% on the AIME math benchmark, comparable to OpenAI's o1, at a fraction of the cost, and reaches a 2,029 Elo rating on Codeforces-style programming challenges
• V3.2 (Dec 2025) — 685 billion parameters, 128K context, released under MIT license with strong reasoning and tool-use capability
• V4 (April 2026) — offered frontier-level performance at very low prices, currently around $0.87 per million output tokens, and can run on Huawei-made processors; the official API ships as two variants — V4 Pro (1.6T total/49B active, quality-focused) and V4 Flash (284B total/13B active, optimized for lower latency)

#Business and Market Position in 2026

DeepSeek's scale has grown substantially:

• 173 million total downloads, 96.9 million monthly active users, and 22.15 million daily active users, with roughly 89% domestic market share in China
• 350.8 million website visits in March 2026 alone
• Third place in enterprise AI market share by 2026, with over 26,000 enterprises integrating its API

In a major shift, DeepSeek raised its first-ever external funding round in 2026 — $7.4 billion at a valuation exceeding $50 billion — bringing in institutional and state-linked investors including Tencent and CATL. This marked a transition from a purely self-funded research lab to a capitalized AI infrastructure company.

#The "Second DeepSeek Shock" (July 2026)

Most recently, the dynamic DeepSeek started has spread to rival Chinese labs. On July 16, 2026, China's Moonshot AI released Kimi K3, the world's first open-weight model with roughly 3 trillion parameters, priced at just $3 per million input tokens — far cheaper than major U.S. frontier models. This echoed the original DeepSeek shock and rattled global investors again, with Asian and U.S. chip stocks selling off the following day. Commentary has noted that what looked like a single anomaly in January 2025 now appears to be a recurring pattern, with capable, cheap, open Chinese frontier models shipping on a regular cadence from multiple labs.

This context is relevant to Claude too — for anything specific to Anthropic models mentioned in these sources (like Fable 5 or Mythos), I'd point you to Anthropic's own statements rather than third-party framing, since coverage of competitors can vary in accuracy.