DeepSeek

startup

china · Founded 2023 · Valuation: $3.4B as of 2025-03

DeepSeek is a dominant force in the open-source AI community, founded in 2023 by Liang Wenfeng, who also founded and leads the quantitative hedge fund High-Flyer (幻方量化). High-Flyer, one of China's largest quant funds managing over $8B in assets, provides DeepSeek's funding and compute infrastructure. The company reportedly operates a cluster of ~10,000 NVIDIA A100 GPUs acquired before US export restrictions took full effect.

DeepSeek triggered a global sensation in January 2025 — the "DeepSeek moment" — when DeepSeek-R1 demonstrated reasoning performance rivaling OpenAI's o1 at a fraction of the training cost. The release briefly wiped nearly $1 trillion from US tech stocks and forced a reassessment of assumptions about AI scaling costs worldwide.

The lab has pioneered several key innovations: Multi-head Latent Attention (MLA) for memory-efficient inference, the DeepSeekMoE architecture with fine-grained expert segmentation, FP8 mixed-precision training at scale, Multi-Token Prediction, and GRPO (Group Relative Policy Optimization) for reinforcement learning on reasoning tasks. DeepSeek is notable for its "open-science" approach, publishing detailed technical reports and open-sourcing low-level training infrastructure including DeepGEMM, FlashMLA, DeepEP, and 3FS (Fire-Flyer File System).

DeepSeek's research was published in Nature in 2025 and the company has become a central case study in debates over AI efficiency, US-China technology competition, and the viability of open-weight frontier models.

Website Wikipedia HuggingFace GitHub Artificial Analysis
frontieropen-weightefficiencymoereasoning

People

News

Outputs (33)

Engram: Conditional Memory via Scalable Lookup

paper

A new axis of sparsity for large language models. Builds upon finding from Geva et al "Transformer Feed-Forward Layers Are Key-Value Memories" (2021).…

mHC: Manifold-Constrained Hyper-Connections

paper

Proposes Manifold-Constrained Hyper-Connections, projecting residual connection space onto a specific manifold to restore identity mapping while maint…

DeepSeek-Math-V2 2

model dataset

Specialized model family for formal and informal mathematical reasoning, with curated training datasets.

↓ 4.9k

LPLB (Linear-Programming Load Balancer)

library

Expert-parallel load balancer using linear programming to optimize MoE workload distribution. Supports Cube, Hypercube, Ring, and Torus topologies. Ea…

DeepSeek-OCR / OCR-2

model

High-efficiency models for optical character recognition and document understanding.

↓ 3M

DeepSeek-V3.2 2

model paper

V3.2 family introducing DeepSeek Sparse Attention for long-context efficiency. V3.2-Speciale achieves gold-medal performance on IMO and IOI 2025, surp…

↓ 291.1k 📄 1

DeepSeek-V3.1

model

V3.1 released 2025-08-21. V3.1-Terminus (max reasoning variant) released 2025-09-22.

DeepSeek-R1-0528

model

Significant upgrade to R1 with enhanced logic and reduced hallucinations.

DeepSeek-Prover-V2

model

RL for subgoal decomposition in formal mathematical reasoning. Includes the DeepSeek-ProverBench evaluation suite.

DeepSeek-GRM: Inference-Time Scaling for Generalist Reward Modeling

paper

Generalist reward model with inference-time scaling.

📄 1

3FS (Fire-Flyer File System)

library

High-performance distributed file system designed for AI training.

Smallpond

library

Lightweight distributed data processing framework built on DuckDB and 3FS. Sorted 110.5 TiB in 30 minutes at 3.66 TiB/min. Released as part of DeepSee…

DualPipe

library

Bidirectional pipeline parallelism algorithm for overlapping computation and communication.

EPLB (Expert Parallelism Load Balancer)

library

Expert-parallel load balancer for DeepSeek-V3/R1 that uses a redundant expert strategy to replicate heavy-loaded experts across GPUs for balanced infe…

DeepGEMM

library

Library for clean and efficient FP8 GEMM kernels with fine-grained scaling.

DeepEP

library

Efficient expert-parallel communication library that bypasses NCCL for MoE communications. Used in DeepSeek models for efficient expert parallelism.

FlashMLA

library

Highly optimized kernels for Multi-head Latent Attention.

NSA: Native Sparse Attention

paper

Hardware-aligned and natively trainable sparse attention mechanism.

📄 2

DeepSeek-R1

model

Incentivizing reasoning capability in LLMs via reinforcement learning. R1-Lite-Preview released 2024-11-20. Full R1 paper 2025-01-20. R1-0528 update r…

↓ 1.6M

DeepSeek-V3 3

model paper

Frontier 671B MoE model with Multi-Token Prediction and FP8 mixed-precision training. V3-0324 update released 2025-03-24. Accompanied by a technical r…

📄 206

DeepSeek-VL2

model

MoE vision-language model for advanced multimodal understanding.

Janus 4

model paper dataset

Unified autoregressive framework handling both multimodal understanding and visual generation (DALL-E style) in one model. Includes Janus and Janus-Pr…

↓ 53.4k 📄 11

DeepSeek-V2.5

model

Combination of V2-0628 and Coder-V2-0724 into a unified model.

↓ 7.2k

Auxiliary-Loss-Free Load Balancing Strategy

paper

A foundational paper for modern Mixture-of-Experts (MoE) architectures that introduces the "Loss-Free Balancing" strategy. It eliminates the tradition…

Fire-Flyer AI-HPC: Cost-Effective Software-Hardware Co-Design

paper

Cost-effective software-hardware co-design for deep learning infrastructure.

DeepSeek-Coder-V2 2

model paper

<p>First open-source MoE code model to beat GPT-4 Turbo on coding benchmarks. The 236B model (21B active) achieved 90.2% on HumanEval, 12.7% on SWE-Be…

↓ 7.4k 📄 48

DeepSeek-Prover

model

Specialized model for formal theorem proving in Lean 4.

↓ 906 📄 5

DeepSeek-V2 2

model paper

Massive 236B MoE model (21B active) that introduced Multi-head Latent Attention (MLA). Accompanied by a technical report.

↓ 12.8k 📄 97

DeepSeek-VL

model

Vision-language model with dynamic tiling encoder for high-resolution image understanding.

↓ 12.7k 📄 43

DeepSeek-Math 2

model paper

Introduced Group Relative Policy Optimization (GRPO) for mathematical reasoning. The foundational model and paper for DeepSeek's math capabilities.

📄 66

DeepSeek-MoE 2

model paper

Pioneering 16B Mixture-of-Experts model with only 2.8B active parameters, setting the stage for DeepSeek's future efficiency focus. Accompanied by a f…

↓ 23.3k 📄 16

DeepSeek-LLM 2

model paper

First general-purpose 67B model, outperforming Llama 2, with a technical report on scaling open-source language models with a long-term vision.

↓ 2k 📄 82

DeepSeek-Coder 2

model paper

Coding-specialized models (1.3B to 33B parameters) trained on 2 trillion tokens, with an accompanying technical report.

↓ 14.2k 📄 104