In a remarkable development within the world of artificial intelligence, the Chinese company DeepSeek has made headlines with the release of its latest model, DeepSeek V3-0324. This newly upgraded language model, which builds upon its predecessor that first appeared in December 2024, has been turned into a powerhouse, surprising many in the industry with its impressive performance across various coding benchmarks.
The release, which occurred quietly on March 24, 2025, came without a grand announcement. However, excitement among users has surged, particularly on platforms like Reddit, where many have begun to hail the model as a formidable competitor to Claude 3.7, especially in the area of code generation. This enhanced version boasts a staggering 685 billion parameters and employs a Mixture-of-Experts (MoE) architecture, indicating a significant leap in technical capability.
According to the benchmarks, DeepSeek V3-0324 achieved remarkable results in various coding tasks. The MMLU-Pro benchmark saw the model's score soar from 75.9 to 81.2, marking a significant improvement of +5.3. Similarly, GPQA showed a stunning increase of 9.3 points from 59.1 to 68.4. The AIME benchmark even hit a new state-of-the-art (SOTA) level, jumping from 39.6 to 59.4, an increase of 19.8 points, revealing a robust advance in performance. On the LiveCodeBench, it also performed well, moving from 39.2 to 49.2 with a boost of 10 points.
DeepSeek’s achievements in the KCORES LLM Arena highlight its impressive standing as the second-best non-reasoning model globally, just behind Claude 3.5 Sonnet, and outperforming records set by Claude 3.7 Sonnet and other rival models developed by OpenAI and Google (including OpenAI's o1 model and Google's Gemini-2.0-Pro-Experimental model). Interestingly, it has managed to outperform Claude 3.7 in several tests, leading to discussions online about the possibility of V3-0324 being trained on Claude 3.7.
Speculation arose as DeepSeek's claims on their performance metrics contrasted notably with Anthropic’s assertions, leading some to wonder if these contentions were politically motivated as accusations of intellectual theft surfaced. While many users were critical of DeepSeek's audacity, the conversation around their performance highlights a mismatch in expectations, particularly regarding their training methodologies.
Among other improvements, DeepSeek V3-0324 enhances frontend web interface development. The model shows remarkable advancements in code writing efficiency, significantly reducing the number of errors in coding. More aesthetically pleasing web pages and gaming interfaces are some of the front-end improvements noted. The Function Calling feature, allowing for the effective invocation of custom tools supplied to the model, has also notably advanced.
Performance feedback from users indicates that the DeepSeek model runs efficiently on high-performance machines, including Mac Studio, achieving approximately 20 tokens per second, which has been initially viewed as a modest pace. However, the achievement of enabling local runs of a model of such magnitude represents a significant technical accomplishment. This indicates a potential for growth in the accessibility of powerful AI tools within the coding community.
As if to underline its grassroots ethos, DeepSeek made its V3-0324 available to the public for free. It is currently accessible through an API and various user-friendly platforms, making it easy for developers and hobbyists alike to test its capabilities.
Beyond performance metrics, the financial implications of using DeepSeek V3 are another significant point of interest. On the Aider polyglot benchmark, it has been reported to have solved coding tasks for around $1, in stark contrast to its competitors, where costs ranged from $15 to $35 for similar tasks. Being cheaper and more efficient adds to its appeal, setting a new standard for cost-effective AI solutions.
In an industry increasingly dominated by proprietary technologies, DeepSeek’s decision to maintain an open-source model under the MIT license is noteworthy. This empowers users not only to engage with the model but also to modify it as per their needs—an attractive proposition compared to more costly, locked-down alternatives like those offered by OpenAI.
As DeepSeek continues releasing models at a rapid pace, from its initial V3 to the R1 reasoning model and now the V3-0324 checkpoint, industry observers see the company as a strong contender against Western AI giants. While it remains premature to declare it a challenger to GPT-4, the momentum it’s gaining among developers and the open-source community is undeniable.
The AI race, particularly between the U.S. and China, is intensifying, and DeepSeek's quiet yet effective strides in innovation deserve keen attention. With each new release, DeepSeek is carving out a significant niche, transforming its position from a newcomer to a serious competitor on the global stage.
