DeepSeek
DeepSeek is a Chinese AI lab whose open-weight models, starting with DeepSeek-V3 (December 2024) and R1 (January 2025), showed frontier-class results at low prices.
Text reviewed October 5, 2026
DeepSeek is a Chinese AI lab whose open-weight models, starting with DeepSeek-V3 (December 2024) and R1 (January 2025), showed frontier-class results at low prices.
Basic
DeepSeek is based in Hangzhou, China. DeepSeek-V3, released in December 2024, is a mixture-of-experts model with 671B total parameters of which 37B are active per token, according to its technical report. DeepSeek-R1, released in January 2025 with open weights, is a reasoning model trained largely with reinforcement learning. Later V4 models continued the series.
Deep
DeepSeek publishes technical reports with architecture and training details, which made its models a reference for the field: mixture-of-experts with many small experts, multi-head latent attention to shrink the KV cache, and FP8 training. R1 showed that reinforcement learning on tasks with checkable answers can produce long chain-of-thought reasoning.
Expert
Because DeepSeek models are open-weight, they are served by DeepSeek's own API and by many third-party hosts at different prices and speeds; BenchGecko tracks each offer. Some deployments restrict or filter topics, which the Gecko Tests Censorship Index measures per model.
DeepSeek keeps shipping new DeepSeek models; the live block on this page lists the newest ones tracked on BenchGecko with their release dates, prices and BenchGecko scores, so this definition never has to name a "latest" model.
Depending on why you're here
- ·A Chinese AI lab with free-to-download models
- ·Compare DeepSeek API prices with third-party hosts for the same weights
- ·See /family/deepseek for every tracked model
- ·Open weights at low prices pressure closed-model pricing
- ·Technical reports detail MoE, multi-head latent attention and FP8 training
- ·R1 (Jan 2025): reasoning from RL on verifiable rewards