DeepSeek-R1 proved that frontier-level reasoning could be trained for a fraction of Western budgets and released under an MIT license. Its launch in January 2025 wiped hundreds of billions off US tech valuations in a day and triggered a global wave of open reasoning models. It remains a cornerstone of the open-weights ecosystem.
DeepSeek no longer serves this model through its own API. The `deepseek-reasoner` name was pointed at DeepSeek-V3.1 on 21 August 2025, then at later generations, and the legacy name was switched off on 24 July 2026; today the vendor's price list carries only the V4 family. The weights stay available under the MIT licence, so the model lives on as a self-hosted download and through third-party providers.