China’s artificial intelligence scene just got a major boost thanks to DeepSeek. The excitement? It’s not just about the upgrade but also the factor used in the AI upgrade. The company on Thursday unveiled an upgraded version of its flagship model, the V3 model, that operates on domestically made chips.
Let’s explore why this is such a big leap for the AI industry in China.
Meet Flagship V3.1.
The original DeepSeek V3 model, released in December 2024, actively trained on 671 billion parameters and featured 128,000-token context windows, enabling it to perform complex tasks with ease. Now, the V3.1 builds on that foundation with some serious enhancements.
Meet DeepSeek V3.1, an AI model with hybrid inference (Think and Non-Think modes). Users can switch between the “non-think” mode that gives standard and direct responses and the “think” mode that offers a more deliberate and multi-step reasoning process.
DeepSeek V3.1 Think offers faster thinking speed compared to DeepSeek-R1-0528. In general benchmarks, the V3.1 scored higher than R1-0528 and V3-0324 at solving coding tasks, multiple languages, and simulated command-line environment assessments.
Although the V3.1 has shorter outputs (15,889 tokens) compared to R1-0528’s 22,615, the V3.1 scored a higher percentage in accuracy and reasoning than R1-0528. With stronger agent skills, the V3.1 is less prone to hallucinations.
When comparing it to the V3 model, which includes a 32K and 128K extension phase, the V3.1 actively increases the 32K extension phase by 10-fold (i.e., 10x more data) and the 128K phase by 3.3-fold. This allows the V3.1 to have a longer memory upgrade and better smarts in math.
The V3.1 actively trained with UE8M0, one of the specific FP8 (8-bit floating point) formats, which lets AI models run faster and cheaper while maintaining accuracy. The UE8M0 makes it compatible with the upcoming domestic chips.
The UE8M0 scalability also allows DeepSeek to train long-context models without running into GPU memory limits.
Why This Matters to China’s Tech
Although China boasts advanced tech and smarts, it is still lags significantly behind the US in the AI race. And the reason for this? China’s dependence on US-made chips like Nvidia’s H800 series, which powered DeepSeek’s V3 model.
The trade war between China and the US, prompted a ban on chip sales to China, which compelling China to find a way to improve their chip-making abilities. And while China has been able to make leeway, it hasn’t been without issues.
For instance, the delay in the release of the DeepSeek R2 model was due to the inability to train the model using Huawei’s Ascend chips. This led to the company resorting back to Nvidia’s chips for training and Huawei’s for inference.
That’s why this move with V3.1 feels like a comeback story. With the UE8M0, the V.31 actively adapts to run on the next generation of domestic chips while offering sustainability, speed, and cost efficiency.
Although the specific chip remains undisclosed, China teases that it may soon unveil hardware that matches or rivals with its imported chips in efficiency.
What’s Next for DeepSeek?
The V.31 actively prepares to leverage domestically made chips, although the vendors remain undisclosed for now. Until Sept 5, 4:00pm UTC, a new pricing system will start off and off-peak discounts for DeepSeek products will end.
DeepSeek’s V.31 release is more than an AI upgrade; it is a bold statement. By ensuring compatibility between its cutting-edge infrastructure and domestic chip compatibility, it has opened a new door to how AI chips will be viewed.
Thanks to this, software developers worldwide might soon see cheaper and more accessible models of AI chips.
