3 weeks ago
14
DeepSeek on Monday announced a new update to its general-purpose AI model DeepSeek-V3. The updated model ‘DeepSeek V3-0324’ now ranks highest in benchmarks among all non-reasoning models.
Artificial Analysis, a platform that benchmarks AI models, stated, “This is the first time an open weights model is the leading non-reasoning model, marking a milestone for open source.” The model scored the highest points among all non-reasoning models on the platform’s ‘Intelligence Index’.
Source: Artificial Analysis
In the GPQA Diamond benchmark, the model achieved a score of 66%, surpassing GPT-4o (54%) and Gemini 2.0 Pro Experimental (62%) and matching Anthropic’s Claude 3.7 Sonnet (66%). This benchmark assesses AI models on complex, graduate-level science questions.
Likewise, the model outperformed all other non-reasoning models across several benchmarks. However, it still trails behind DeepSeek-R1, OpenAI’s o1, o3-mini, and other reasoning models.
Reasoning models consume additional time to perform a step-by-step thinking process to respond, whereas non-reasoning models prioritise speed and often respond immediately.
Source: Artificial Analysis
The performance of DeepSeek V3-0324 across all popular benchmarks can be found in Artificial Analysis.
It is also rumoured on X that DeepSeek V3-0324 may be the base model for the forthcoming DeepSeek-R2 reasoning model. Recently, Reuters reported that DeepSeek plans to release R2 “as early as possible”. The company initially intended to launch it in early May but is now contemplating an earlier timeline.
The model is expected to produce “better coding” and can reason in languages beyond English. “This release is arguably even more impressive than R1—and potentially indicates that R2 is going to be another significant leap forward,” added Artificial Analysis.
A few months ago, DeepSeek shook the AI ecosystem and significantly impacted NVIDIA’s market cap by providing state-of-the-art performance despite using a minimal number of GPUs for training.
In addition to their impressive performance, models from DeepSeek are also favoured for their cost efficiency. It was recently announced that DeepSeek would provide discounts for its API platform during non-peak hours – from 16:30 to 00:30 daily.
In a recent GitHub post, the company reported a theoretical daily profit margin of 545% for its inference services despite the limitations in monetisation and discounted pricing structures.
While Chinese AI models rival those from the United States, fierce competition exists among the major players within China too. Big-tech companies like Alibaba, Baidu, Tencent, and ByteDance have all been regularly announcing AI models across multiple domains, each trying to outperform the other.
