Skip to content
Search

DeepSeek

lastmod 2026-05-13created 2024-12-3122 edits

Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., trading as DeepSeek, is a Chinese artificial intelligence company, based in Hangzhou, Zhejiang, that develops open-source large language models (LLMs). It is owned and funded by Chinese hedge fund High-Flyer, and was founded by Liang Wenfeng, who established the company in 2023 and serves as its CEO.

DeepSeek v3 vs GPT-4o vs 3.5 Sonnet

Final verdict
Reasoning: Deepseek v3 > Claude 3.5 Sonnet > GPT-4o
Math: Deepseek v3 > Claude 3.5 Sonnet > GPT-4o
Coding: Claude 3.5 Sonnet > Deepseek v3 ~ GPT-4o
Creative writing: Claude 3.5 Sonnet > Deepseek v3 ~ GPT-4o
Who should use Deepseek v3?
Can replace GPT-4o for most tasks
Optimal if you are building applications. Excellent price-performance makes it well suited for developing user applications
Open-weight, so it can be self-hosted, giving greater control

DeepSeek-V3.2

  • DeepSeek-V3.2 released - pushing the limits of open large language models | GeekNews
  • DeepSeek-V3.2 is an open-source large language model combining high computational efficiency with reasoning and agent performance
  • The new DeepSeek Sparse Attention (DSA) architecture maintains performance even in long contexts while greatly reducing computational complexity
  • Achieves GPT-5-level performance through a scalable reinforcement learning (RL) framework; the high-performance version attains reasoning ability on par with Gemini-3.0-Pro
  • A large-scale agentic task synthesis pipeline generated 1,800 environments and 85,000 prompts, improving generalization and instruction following in complex interactive environments
  • Open models narrow the gap with closed models and establish themselves as a cost-effective alternative

Why DeepSeek is cheap at scale but expensive locally

Some AI models such as DeepSeek-V3 are cheap and fast when served at scale but slow and expensive when run locally.

  • The reason lies in the fundamental trade-off between throughput (throughput) and latency (latency) related to GPU utilization efficiency.
  • Increasing the batch size makes the GPU work efficiently, but users must wait until tokens accumulate, which increases latency
  • Models with a Mixture-of-Experts architecture and deep pipelines require large batches and high latency
  • In a local single-user environment it is hard to form large enough batches, leading to degraded performance and higher costs
  • OpenAI, Anthropic, and others achieve fast responses through more efficient architectures, sophisticated batching strategies, or throwing excessive GPUs at the problem

See also

  • LLM
  • DeepScaleR - surpassing O1-Preview with a 1.5B model using RL
  • DeepEP - an efficient expert-parallel communication library
  • FlashMLA - efficient MLA decoding kernels for Hopper GPUs
  • DeepGEMM - a library for efficient FP8 matrix multiplication (GEMM) that supports the fine-grained scaling (fine-grained scaling) approach proposed in DeepSeek-V3
  • 3FS - DeepSeek's distributed file system
  • DeepSeek OCR
  • DeepSeek Math
  • GLM-5 - a model targeting complex systems engineering and long-horizon agent tasks
  • DeepSeek-TUI - a coding agent for DeepSeek models that runs in the terminal

Favorite site

author Zer0 sitemap