DeepSeek R2: the Chinese lab relaunches its open-weight reasoning model
In one sentence DeepSeek ships R2, successor to R1: more efficient step-by-step reasoning, open weights, contained training cost. Fresh pressure on closed reasoning models.
Exactly one year after R1 (January 2025), DeepSeek releases R2: the successor to the model that in early 2025 shook the market by proving high-quality reasoning was achievable on a training budget far below American players.
R2 repeats the formula: open weights on HuggingFace, permissive license, focus on step-by-step (chain-of-thought) reasoning trained via reinforcement learning. The difference is a full year of accumulated improvements: refined MoE architecture, better data, more aggressive distillation to smaller models.
For self-hosters: the smaller version runs on a single high-end consumer GPU, and quantized builds for Ollama / llama.cpp will follow within days.
Politics and market: China shows the frontier-reasoning chase didn't stop at US export controls. The "sovereign European AI" debate gains another argument.
Companies
DeepSeek
Tools
DeepSeek R2
Tags
Sources