DeepSeek

DeepSeek is a Chinese AI research lab that develops and open-weights large language models, best known for triggering a major shift in AI cost expectations when it released DeepSeek-R1 in January 2025 — a reasoning model competitive with leading Western models at a fraction of the reported training cost, which briefly wiped out hundreds of billions of dollars in US tech stock value on release. The company (formally Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.) was founded in July 2023 by Liang Wenfeng, who also co-founded and runs High-Flyer, the Chinese quantitative hedge fund that funded DeepSeek entirely on its own until April 2026, when DeepSeek raised roughly $7.4 billion (about 50 billion yuan) in its first external funding round; reported valuations for that round vary by source, generally landing somewhere in the tens of billions of dollars.

DeepSeek's current flagship is the V4 family, which reached general availability on August 13, 2026: V4-Pro (1.6 trillion total parameters, about 49 billion active per token, built for coding and complex agent tasks) and V4-Flash (284 billion total parameters, about 13 billion active, tuned for speed and lower cost). Both share a 1 million-token context window and a hybrid attention architecture, a departure from the Multi-head Latent Attention used in V2 and V3. DeepSeek also maintains DeepThink (R1), a reasoning model that shows its step-by-step chain of thought before answering. Every recent flagship release — V4-Pro, V4-Flash, V3.2, V3.1, and R1 — is published as open-weight under the MIT license on Hugging Face, so anyone can download and run the models locally with tools like Ollama or vLLM, in addition to using DeepSeek's free chat app or its token-billed API.

1. Chat for free at chat.deepseek.com or through the DeepSeek mobile app — no payment required for standard chat use.
2. Choose a mode for your task: standard chat uses the V4 family by default, while DeepThink (R1) can be toggled on for problems that benefit from visible step-by-step reasoning, such as math, logic, or code.
3. For programmatic access, sign up for the DeepSeek API and call V4-Pro or V4-Flash by model name, paying per token, with lower rates for cached input and off-peak hours.
4. To run the models yourself instead of using DeepSeek's hosted service, download the open MIT-licensed weights for V4-Pro, V4-Flash, V3.2, V3.1, or R1 from Hugging Face and serve them locally or on your own infrastructure with a tool like Ollama or vLLM.
5. Be mindful of DeepSeek's peak/off-peak API pricing, introduced August 16, 2026: rates roughly double during peak windows (01:00-04:00 and 06:00-10:00 UTC), so batching non-urgent API workloads outside those hours can meaningfully cut cost.

• Free, general-purpose chat and reasoning assistance through the web or mobile app, without a subscription.
• Coding and complex agentic tasks using V4-Pro, the larger of the two V4 variants.
• Cost-sensitive, high-volume API workloads using V4-Flash, priced well below V4-Pro and most Western frontier models.
• Problems that benefit from visible chain-of-thought reasoning — math, logic puzzles, multi-step code generation — using DeepThink (R1).
• Self-hosting an open-weight frontier-class model on private infrastructure, for organizations that can't send data to a third-party API, using the MIT-licensed weights on Hugging Face.
• Researchers and developers building on or fine-tuning open model weights rather than starting from scratch, given DeepSeek's consistent MIT licensing across releases.