DeepSeek

Liang Wenfeng founded DeepSeek in Hangzhou to develop large AI models. Its origins were connected with the technical work of High-Flyer, which had previously invested in GPU clusters.⁠[4]

The team published research on scaling 7-billion- and 67-billion-parameter DeepSeek LLM models. It described pretraining on two trillion tokens, followed by supervised fine-tuning and direct preference optimization to produce chat models.⁠[5]

DeepSeekMoE research divided expert networks more finely and introduced shared experts for common knowledge, reducing redundancy among routed experts. The approach aimed to expand total capacity while controlling computation per use.⁠[6]

DeepSeekMath continued training a code model with mathematical material and introduced Group Relative Policy Optimization, or GRPO. The reinforcement-learning method sought to improve mathematical reasoning while reducing memory requirements relative to conventional PPO training.⁠[7]

The V2 report described a mixture-of-experts model with 236 billion total parameters, 21 billion activated per token and a 128K context. Multi-head Latent Attention compresses the key-value cache, complementing sparse expert computation to reduce inference resource requirements.⁠[8]

V3 used 671 billion total parameters with 37 billion activated per token, auxiliary-loss-free load balancing and a multi-token prediction objective. The team reported 2.788 million H800 GPU hours for training; that compute figure is not the company’s total research expenditure.⁠[9]

R1 research demonstrated how reinforcement learning could encourage reasoning behaviors including self-checking, verification and strategy adaptation. It also examined transferring reasoning patterns from large models to smaller ones, with evaluations centered on verifiable mathematics, coding and science tasks.⁠[10]

R1-0528 made model weights available and added JSON output and function calling. DeepSeek described improvements in front-end tasks and fewer hallucinations; those release claims did not guarantee the accuracy of every generated response.⁠[11]

V3.1 combined thinking and non-thinking modes in one model, both with 128K context. The release added AnthropicAnthropicAnthropic is an artificial intelligence research and product company founded in 2021 that develops the Claude models, assistant, and tools such as Claude Code, and operates as a public benefit corporation. Dario Amodei and Daniela Amodei are its co-founders, and its research covers model safety, interpretability, and steerability.Open the full entry-format API compatibility and strict function calling in the beta API, emphasizing tool use and multistep tasks.⁠[12]

The experimental V3.2-Exp introduced DeepSeek Sparse Attention on top of V3.1-Terminus. Selective attention aimed to reduce long-context training and inference computation, with the model offered through web, app and API services.⁠[13]

DeepSeekMath-V2 combined proof generation with verification, encouraging models to inspect and repair proofs before finalizing them. The research emphasized that a correct final answer does not ensure rigorous reasoning, extending training toward completeness and reliability of proofs.⁠[14]

V3.2 integrated thinking with tool use alongside Speciale, a variant emphasizing complex reasoning. Speciale was offered through a temporary evaluation API without tool calls, scheduled to end on December 15; that announcement did not establish a permanent endpoint.⁠[15]

Engram research treated knowledge lookup as a capability complementary to neural computation, using scalable lookup tables for conditional memory. It explored delegating common patterns to memory modules so that the main network could devote more computation to complex reasoning.⁠[16]

DeepSeek-OCR 2 explored dynamically reordering visual tokens according to image semantics rather than a fixed top-left-to-bottom-right scan. It proposed an encoder with causal reasoning for complex layouts and made code and model weights available.⁠[17]

The privacy policy described collecting inputs, uploaded files, chat history and device information, with personal data processed and stored in China. It also described model-improvement uses and channels for exercising applicable access, deletion and training opt-out rights.⁠[18]

Anthropic accused DeepSeek of extracting Claude capabilities through over 150,000 exchanges involving reasoning and grading tasks. This was the company’s allegation.⁠[19]

The V4 preview offered Pro and Flash variants, both with million-token context and thinking or non-thinking modes. Pro had 1.6 trillion total and 49 billion active parameters; Flash had 284 billion total and 13 billion active parameters, serving different compute and responsiveness requirements.⁠[20]

V4.1-Flash introduced native visual understanding in a 552-billion-parameter mixture-of-experts model. Its causal encoder–decoder design activated eight billion parameters for input and 16 billion for output, using different computation configurations for input processing and generation.⁠[21]

V4.1-Flash was upgraded to version 1017 across the official web, app and API services, keeping the calling name and pricing unchanged at that update. DeepSeek reported improved coding and agent-task performance and opened access to its in-house search engine.⁠[22]

Search returns structured search results, while Fetch converts an individual web page into Markdown for agents or retrieval-augmented applications. Both use the same API authentication as the generative models.⁠[23]