DeepSeek

DeepSeek (深度求索) is a Chinese artificial-intelligence research lab founded in Hangzhou in May 2023 by Liang Wenfeng, funded by his quantitative hedge fund High-Flyer. Small (~150 people), flat, and explicitly organized around an AGI goal rather than a product roadmap, DeepSeek pairs fully open weights (MIT license) with an aggressively cheap per-token API. Its stated strategy is restraint: price to recoup hardware costs in roughly ten months, refuse consumer/enterprise monetization fights, and treat both as “side products on the road to AGI.” ¹

Model lineage

Release	Params (total/active)	Key advances
V2 (2024-05)	236B / 21B	Multi-head Latent Attention (MLA); DeepSeekMoE fine-grained experts
V3 (2024-12)	671B / 37B	Auxiliary-loss-free load balancing; multi-token prediction; FP8 training + DualPipe; ~2.788M H800-hours total
R1 (2025-01)	671B / 37B	RL-first reasoning via GRPO (critic-free group-relative advantages)
V3.1 → V3.2 (2025)	671B / 37B	Hybrid thinking/non-thinking single model; agent-focused post-training
V4 Preview (2026-04-24)	1.6T / 49B (Pro); 284B / 13B (Flash)	CSA+HCA sparse attention; mHC; Muon optimizer; 1M-token context default; MIT license
V4-Flash GA “0731” (2026-07-31)	284B / 13B	Same architecture, re-post-trained only; beats V4-Pro-Preview on agentic benchmarks

The legacy deepseek-chat/deepseek-reasoner API aliases were hard-retired 2026-07-24; the V4-Pro GA remains pending (“will follow soon”) as of early August 2026. ²

Technical through-line

DeepSeek’s architecture is a continuous refinement of one bet: extreme MoE sparsity + aggressive attention compression + training efficiency, in service of frontier-adjacent capability at minimal compute. The V4 headline is a tiered sparse-attention stack replacing dense attention: HCA (Heavily Compressed Attention) compresses keys ~128× and runs dense attention on the coarse representation in early layers; CSA (Compressed Sparse Attention) compresses ~4× and applies DeepSeek Sparse Attention (DSA) — a lightweight “lightning indexer” scores compressed blocks and selects top-k tokens (1024 Pro / 512 Flash) for full attention in later layers. At 1M-token context, V4-Pro needs only ~27% of the single-token inference FLOPs of V3.2 (Flash ~10%), and mixed-precision KV storage (RoPE dims BF16, rest FP8) roughly halves the cache again — which is what makes a million-token context the default tier rather than a premium one. Supporting pieces: mHC (Manifold-Constrained Hyper-Connections) — residual streams widened 4× with the mixing matrix constrained to the doubly-stochastic manifold via Sinkhorn iterations — and the Muon optimizer, under which DeepSeek reports zero loss spikes across the ~32T-token pre-training run. ³

Business and compute position

• Pricing as doctrine. V4-Pro launched 2026-04-24 at $1.74/$3.48 per M tokens, cut ~75% to $0.435/$0.87 within five weeks; Flash sits at $0.14/$0.28 with cache-hit input orders of magnitude lower. Liang’s pricing rule is hardware recoup in ~10 months, explicitly not profit maximization; a peak/off-peak surge model (2× Beijing business hours) is announced but not yet active. ⁴
• Funding. June 2026 brought the first-ever external round (~$7.4BfromTencent,CATL,andothersata$50B valuation), with a second round and onshore IPO rumored — a research lab becoming a real company on Chinese strategic-capital terms.
• Ascend pivot. V4 is optimized for Huawei Ascend silicon; Liang frames the China–US gap as “mainly resources… on people, almost no gap,” chasing with roughly 1/20th the compute, and treats domestic-chip adoption (Huawei 950 supernodes) as both an export-control hedge and an ecosystem investment. ⁵
• Reality check. V4 is not the absolute capability leader — it trails the top US closed models on composite evals — and inside China it is the price floor, not the quality leader, with Moonshot, Zhipu, Alibaba, MiniMax and others pressing. Liang predicts consolidation: “There are too many base-model companies in China; it will converge.”

Philosophy (leaked investor call, July 2026)

A leaked four-hour Liang Wenfeng investor call (talk dated 2026-05-20, public 2026-07-23; two independent English translations) is the richest statement of DeepSeek’s self-understanding: vision over organization (“we don’t really have an organization… that vision isn’t even written down”), open source as the point rather than the tactic (“I want [competitors] to be able to deploy it. I only worry they can’t”), restraint as strategy (“People who take more get beaten by people who take less”), and a roadmap of CoT → agents → continual learning → a gradual “singularity” → embodied intelligence — while refusing video, 3D, and world models as off the “main line of intelligence.” The call is internally consistent with every observable decision (price cuts, MIT license, tiny team), though authenticity is corroborated rather than officially confirmed. ⁶

Connections

• chinese-innovation-patent-ecosystem: R1’s January 2025 release was the archetypal Chinese “Sputnik moment” in the patent-literature sense; DeepSeek is the firm-level proof point for that page’s measurement debates.
• gongkai: DeepSeek’s open-weights, ecosystem-first strategy is gongkai’s quid-pro-quo logic in AI form — give away the artifacts, own the ecosystem position. Liang’s “open source is a cultural behavior” stance echoes the network-IP view bunnie Huang documented in Shenzhen hardware.
• technological-singularity: Liang’s “singularity” is explicitly gradual and continual-learning-gated — a concrete lab-level operationalization of a usually abstract concept.
• paperclip-maximizer: An AGI-aiming lab whose stated philosophy is taking less — a live counterpoint to pure instrumental-convergence pessimism, worth tracking as it scales.
• specification-gaming-openai-hf-incident / anthropic-cybersecurity-eval-incidents: The frontier-lab cluster DeepSeek now belongs to structurally, whether or not it shares the incidents.

Caveats

• The investor-call transcript is a leak (widely corroborated, not officially confirmed); funding/IPO details are partly rumor; benchmark numbers mix official self-reports and third-party evals.

Sources

• 2026
• 2026 — Change Log
• 2026 — DeepSeek founder Liang Wenfeng in His Own Words: 64 Quotes from DeepSeek’s Investor Call

See Also

• llm-inference-provider-landscape — DeepSeek’s first-party API is the archetypal raccoon-free fallback tier (creator + hoster in one)
• chinese-innovation-patent-ecosystem
• gongkai
• technological-singularity
• paperclip-maximizer
• specification-gaming-openai-hf-incident

Footnotes

1. 2026 — DeepSeek founder Liang Wenfeng in His Own Words: 64 Quotes from DeepSeek’s Investor Call ↩

2. 2026 — Change Log ↩

3. 2026 ↩

4. 2026 — DeepSeek founder Liang Wenfeng in His Own Words: 64 Quotes from DeepSeek’s Investor Call ↩

5. 2026 — DeepSeek founder Liang Wenfeng in His Own Words: 64 Quotes from DeepSeek’s Investor Call ↩

6. 2026 — DeepSeek founder Liang Wenfeng in His Own Words: 64 Quotes from DeepSeek’s Investor Call ↩