Vibeleaderboard
DeepSeek

ai lab

Also indexed as deep-seek · deepseek-ai

DeepSeek

DeepSeek matters because it paired competitive open-weight models with documented techniques for lowering training and inference costs, then made a strong reasoning recipe available for study and adaptation. Its impact is technical and economic: it changed what practitioners expected from Chinese open models and forced closer scrutiny of the assumed relationship between compute spending, model access, and capability.1,2,3,8

Official site ↗Models in the Atlas →DeepSeek on X ↗DeepSeek on GitHub ↗
Overall Viber #25

14 tools · 3 Intel

Profile

Overview

Origins in High-Flyer

DeepSeek is a Hangzhou AI laboratory founded in 2023 by Liang Wenfeng. It grew from the research and computing base of High-Flyer, the quantitative investment firm Liang co-founded. Independent profiles describe High-Flyer as DeepSeek's financial backer, while DeepSeek's papers identify Liang as a model author. The relationship helps explain the lab's early access to large compute clusters, but the company publishes little conventional corporate history.6,7

An efficiency-centered model program

The lab first established a broad technical program through DeepSeek LLM, Coder, Math, and VL. DeepSeek-V2 then introduced Multi-head Latent Attention and combined it with DeepSeekMoE to reduce inference memory and computation. DeepSeek-V3 extended that efficiency program with FP8 training, auxiliary-loss-free load balancing, and multi-token prediction, while publishing open weights and a detailed technical report.1,2

R1 and open reasoning research

DeepSeek-R1 moved the organization from architectural influence to global attention. Its report documented R1-Zero, trained with large-scale reinforcement learning before supervised fine-tuning, and the production R1 model, which added cold-start data and further training stages. The lab released R1, R1-Zero, and distilled models based on Qwen and Llama. The result gave researchers and deployers inspectable reasoning models, though open weights do not make the undisclosed training data fully open source.3,8

From price disruption to service economics

DeepSeek operates a consumer assistant and low-cost API alongside its research releases. Its 2026 transparency center describes model behavior and compliance practices, while Reuters reported large increases and time-dependent pricing for V4-Pro and V4-Flash. The factual shift is from introductory shock pricing toward segmented serving economics. The editorial uncertainty is whether that pricing supports a durable service without weakening the open ecosystem that produced much of DeepSeek's influence.5,9

Company evidence

Epoch AI dataset ↗
Reported revenue
$450MAnnualized run rate
Jul 14, 2026 · Likely[1]
Latest funding
$7.5B$50B post-money valuation
Jun 16, 2026 · Likely[1][2]
Reported staff
300Full company
Jun 26, 2026 · Likely[1]
Reported usage
10M daily usersDeepSeek
Mar 28, 2025 · Likely[1][2]

Reported estimates, not audited figures. Confidence labels and source links are preserved from the dataset.

Reporting and context

Notable contributions

  1. 01Multi-head Latent AttentionDeepSeek-V2 introduced Multi-head Latent Attention, which compresses key and value representations to reduce the cache required during inference while retaining model quality.1
  2. 02Group Relative Policy OptimizationDeepSeekMath introduced Group Relative Policy Optimization as a reinforcement-learning method that estimates advantages from groups of sampled outputs without a separate critic model. DeepSeek later used GRPO in the R1 program.4,3
  3. 03Open reinforcement-learning reasoning artifactsThe R1 release paired weights with a report on R1-Zero, cold-start data, reinforcement learning, and distillation, giving outside researchers unusually concrete artifacts for studying reasoning behavior.3
  4. 04FP8 training at large scaleDeepSeek-V3 documented successful FP8 mixed-precision training for a 671 billion parameter mixture-of-experts model, connecting numerical efficiency to a public frontier-scale training report.2