Chinese AI company DeepSeek has released an updated version of its open-source reasoning model. Called DeepSeek-V2-R1+, the new model can now handle much longer inputs—up to 128,000 tokens at once—and promises better results in math, coding, and logic tasks.
The original R1 model was launched in April 2024. This new version builds on it using a “Mixture of Experts” (MoE) design. In simple terms, the model doesn’t use all of its parts at once—it activates only what’s needed, saving on computing power without reducing quality. This method is also used by other big AI labs like Google DeepMind and Mistral.
Model scores improve across key tests
According to DeepSeek, the new R1+ model shows better scores across several standard AI tests. These include:
- MATH: 81.3
- GSM8K (grade-school math): 80.4
- HumanEval (code writing): 83.9
- GPQA (graduate-level questions): 92.1
These results show small but steady improvements over the earlier version. While it doesn’t beat top AI models like OpenAI’s GPT-4 or Google’s Gemini, it holds its own among open-source models.
The extended context window is also a big step. It lets the model handle longer conversations, summarize lengthy documents, and solve problems with many steps—tasks that shorter-context models struggle with.
Part of China’s growing open-source AI movement
DeepSeek is one of several Chinese companies working on open-source AI tools. Others include Baichuan, InternLM, and Moonshot AI. By sharing their models publicly, these companies want to give researchers and developers more flexible options compared to closed, commercial tools.
China’s open-source push is also seen as a way to compete globally in AI innovation, especially as access to Western technologies becomes more limited.
Still behind the top global models
Even with this update, R1+ is not at the same level as the top-performing commercial models like GPT-4 or Claude 3. It does well in focused reasoning tasks, but overall capabilities still lag behind.
DeepSeek has not shared all technical details, like training data or compute power used. But the release shows that Chinese labs are continuing to make steady progress—and are serious about staying in the global AI race.
