DeepSeek

DeepSeek is a family of large language models developed by the Chinese AI lab DeepSeek (Hangzhou DeepSeek Artificial Intelligence). Since early 2024 the company has released a rapid succession of frontier-class models, all published as open weights under the MIT license — making them freely downloadable, fine-tuneable, and deployable without royalties.

Model families

• DeepSeek-V2 (May 2024): Introduced the Mixture-of-Experts (MoE) + Multi-Head Latent Attention (MLA) architecture that underpins the entire modern DeepSeek lineage. 236B total parameters with 21B active per token; 128K context.
• DeepSeek-V3 (December 2024): Flagship general-purpose model. 671B total parameters with ~37B active per token (MoE). Matches or beats GPT-4o and Claude 3.5 Sonnet on coding, reasoning, and knowledge benchmarks. 128K context window. ~90K GitHub stars on the DeepSeek-V3 repo.
• DeepSeek-R1 (January 2025): Reasoning-specialised model built on the V3 backbone, trained with large-scale reinforcement learning to produce extended chain-of-thought (CoT) outputs. Matches OpenAI o1 on AIME, MATH-500, and code competition benchmarks. 671B params; 128K context. Released alongside six distilled variants (1.5B–70B) based on Qwen and Llama. DeepSeek-R1 repo reached ~92K GitHub stars within weeks of release.
• DeepSeek-Coder / DeepSeek-Coder-V2 (June 2024): Code-specialised model. V2 extends context from 16K to 128K and uses a 236B MoE base; outperforms GPT-4 Turbo on HumanEval and DS-1000 coding benchmarks.
• DeepSeek-V3.1 / V3.2 (Aug–Dec 2025): Iterative updates adding hybrid thinking/non-thinking modes and extending context to 164K tokens (V3.2).
• DeepSeek-V4 (April 2026, preview): Two-variant release — V4-Pro (1.6T total / 49B active params) and V4-Flash (284B total / 13B active), both with 1M token context windows.

Access methods

Self-hosting: Weights for all generations are published on Hugging Face under deepseek-ai/. Models can be served locally via Ollama (ollama run deepseek-v3), vLLM (vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B), or LM Studio. Smaller distilled variants (7B–32B) run on consumer hardware; full V3/R1 (671B) requires a multi-GPU server (e.g. 8× H100).

DeepSeek API: platform.deepseek.com provides an OpenAI-compatible REST API. Pricing is dramatically cheaper than comparable closed-source models — DeepSeek V3 costs ~$0.27/$1.10 per million input/output tokens vs GPT-4o at $2.50/$10.00 (≈9× cheaper). DeepSeek R1 costs $0.55/$2.19 per million tokens vs OpenAI o3 at $150/$600 (>200× cheaper). Cache hits reduce input costs by 90%. The API is hosted primarily in China and can experience capacity constraints during peak demand.

Third-party providers: DeepSeek R1 and V3 are available on Amazon Bedrock (fully managed, serverless), Fireworks AI, Together AI, and OpenRouter — providing enterprise-grade reliability with global data residency options.

License: MIT for V3, R1, Coder-V2, and all distilled variants — fully permissive for commercial use, fine-tuning, and redistribution.

Why it matters: DeepSeek demonstrated that open-weight models trained with efficient MoE architectures can match closed-source frontier models at a fraction of the compute cost. The January 2025 R1 release triggered a significant re-evaluation of AI development economics across the industry.