DeepSeek

Hangzhou-based open-weight LLM lab funded by a quant hedge fund — became the symbol of cost-efficient frontier AI after January 2025.

1. Core Product / Service

DeepSeek builds open-weight large language models with a Mixture-of-Experts (MoE) architecture. Public-facing surfaces:

• DeepSeek API — chat + reasoning endpoints, OpenAI-compatible.
• Open-weight releases on Hugging Face (V3, V3.1, V3.2, R1, V4 Flash, V4 Pro), code under MIT, weights under DeepSeek's own model license [1][2].
• Web/app chat at chat.deepseek.com.

Latest releases (as of 2026-05-09):

• V4 Pro (preview, 2026-04-24): 1.6T total / 49B active params, 1M context. Claims ~27% of V3.2's per-token inference FLOPs and ~10% of its KV cache at 1M context [3][4].
• V4 Flash: 284B total / 13B active, 1M context [3].
• R1: reasoning model line, originally released 2025-01-20 [5].

2. Target Users & Pain Points

• Developers / startups wanting frontier-tier reasoning at a fraction of OpenAI/Anthropic prices.
• Self-hosters / sovereign AI — open weights mean they can run on own GPUs (vLLM/SGLang Day-0 support; see ai-inference-engines).
• Researchers — published technical reports for V3 and R1 are unusually detailed for a frontier lab.

Pain points addressed: closed-model lock-in, US-API egress cost, training-cost transparency.

3. Competitive Landscape

Differentiation: DeepSeek is the most aggressive on inference cost per quality unit and on architecture novelty (MoE routing, multi-head latent attention). Strategic choice to upstream optimizations into ai-inference-engines (vLLM) rather than build their own engine — keeps team focused on model training (Internal research notes).

4. Unique Observations

• Inference-engine strategy: DeepSeek explicitly chose to merge optimizations back into vLLM and SGLang rather than launch a competing inference product. Likely reasoning: a ~500-person team can't both train new models and maintain a multi-hardware inference stack — let the engine ecosystem distribute their models for them (Internal research notes).
• Search stack: independent Chinese labs (DeepSeek, Kimi, MiniMax) reportedly mix self-built crawl with Exa.ai for grounding, unlike Baidu/Alibaba/Tencent which use in-house engines (Internal research notes).
• OpenRouter behavior quirk: DeepSeek models served via openrouter can have reasoning enabled by the host, while direct DeepSeek API defaults reasoning off — relevant for hermes-openrouter-models routing.
• V4 inference-FLOP claim: V4 Pro's 27%-of-V3.2 single-token compute at 1M context is the headline architectural improvement, not raw benchmark scores [3][4].

5. Financials / Funding

• No external venture funding. DeepSeek has not raised a public VC round; venture firms initially passed because there was no near-term exit [6].

• Parent / funder: High-Flyer (幻方量化), a Chinese quantitative hedge fund founded 2016 in Hangzhou by Liang Wenfeng + Zhejiang University classmates. High-Flyer subsidizes DeepSeek's GPU clusters and operating costs [6][7].

• Ownership: as of May 2024, Liang Wenfeng personally held ~84% of DeepSeek via two shell entities [6].

• Disclosed training cost: V3 base model trained for ~$5.576M in H800 rental-equivalent (2,048× H800, ~55 days), plus ~$294K for the R1 RL phase — figures cover compute only, not salaries / data / failed runs / hardware capex [8]. The $5.6M number is the one that triggered the market reaction below.

• Market impact event (2025-01-27): R1 release on 2025-01-20 plus the cheap-training narrative drove a ~17% Nvidia drop, wiping ~$589B in market cap in a single day — the largest single-day loss in US stock-market history at the time. Broadcom (-17%), Micron (-12%), AMD (-6%) sold off in sympathy [9]. Nvidia recovered ~76% from that low by April 2026.

• Pricing convergence (May 2026): DeepSeek and MiniMax/Mimo pricing have converged to essentially the same band, eroding DeepSeek's earlier absolute cost advantage. The Chinese frontier LLM market is approaching commodity pricing, with differentiation shifting from price-per-token to architectural features (context length, agent capability, multimodal breadth) (Internal research notes).

• GLM 5.2 enters the tier (June 2026): zhipu's GLM 5.2 upgrade now directly competes with DeepSeek-V4-Pro on Chinese domestic benchmarks, further compressing differentiation. With GLM and MiniMax both at parity with DeepSeek on core reasoning/coding tasks, the Chinese frontier market is a four-way contest between DeepSeek, GLM, kimi, and MiniMax — all within a narrow performance band and converging price envelope. The moat is shifting to vertical specialization (GLM's GUI agent, Kimi's long-context, MiniMax's multimodal) and distribution (OpenRouter ranking, enterprise channel) (Internal research notes).

• Price increase (Aug 2026): DeepSeek raised API prices on its V4 family in early August 2026, drawing mixed community reaction to the V4 release + repricing and prompting questions about knock-on pricing at MiniMax/MiMo and kimi (unresolved as of Aug 2026). Specific new rates were not captured (Internal research notes).

6. People & Relationships

• Founder / CEO: Liang Wenfeng (梁文锋) — also CEO of High-Flyer, Zhejiang University grad, started algorithmic trading during the 2008 financial crisis.
• Parent: High-Flyer hedge fund (Hangzhou, founded 2016).
• Ecosystem allies: vLLM / ai-inference-engines (DeepSeek upstreams MoE / expert-parallel optimizations there); SGLang (Day-0 support for V3/R1).
• Distribution partners: openrouter (third-party host, often cheaper than direct API), DeepInfra, Together.
• Peer Chinese labs: kimi (Moonshot), MiniMax, Qwen — overlapping but differentiated positioning (see section 3).