based on materials from the site - By Cryptopolitan_News
The Chinese startup DeepSeek has introduced a new update, claiming that it outperforms the widely known R1 in key metrics. On Thursday, in WeChat, the AI development company confirmed that the new version of the model, V3.1, provides faster responses to queries and signals the start of AI agent development.
DeepSeek added that the model supports a hybrid thinking architecture, which has both thinking modes and non-thinking modes, enhanced agent capabilities, and higher performance when using tools and performing tasks.
DeepSeek provides a 'Deep Thinking' button to switch between modes.
The official DeepSeek app and website have already been updated to version 3.1, allowing users to switch between thinking and non-thinking modes using the 'Deep Thinking' button, similar to how hybrid models from Anthropic, such as Opus and Sonnet, operate.
The V3.1 model is reported to also show better results in benchmarks such as SWE and Terminal-Bench, as well as higher thinking efficiency than R1. Moreover, according to data from Artificial Analysis, the intelligence index of the model in reasoning mode reached 60 points, slightly above the 59 points scored by R1. The underlying architecture remains the same: 671 billion total parameters and 37 billion active.
Despite the higher efficiency, in reasoning mode it uses slightly fewer tokens than R1. However, the new model slightly lags behind the latest Alibaba model and the open-source reasoning model OpenAI GPT-OSS in performance. Additionally, it lacks function calling in reasoning mode, which is considered a serious limitation in agent workflows.
The startup first announced the new model on Tuesday, although at that time it was only available on Hugging Face. In a separate statement, it was added that the version was adapted to work on next-generation AI chips produced in China.
Now, the company has introduced a new pricing plan for its updated version V3. The plan includes an increase in some rates, the cancellation of evening discounts, and a reduction in costs for some applications starting September 6.
DeepSeek has set the price for its Input API at $0.07 per million tokens in cache hits and $0.56 for misses in cache, while for output tokens it is $1.68 per million. These prices are significantly lower than competitors: Gemini 2.5 Pro costs $10 per million output tokens ($15 for longer requests), OpenAI's GPT-5 also costs $10, and Claude Opus 4.1 from Anthropic is $75.
Analysts expected DeepSeek to release a successor to R1 early this year.
DeepSeek first shook Silicon Valley by launching its affordable and powerful AI model R1 in January. Since then, this model has remained at the forefront of the accelerating development of AI in China, challenging American companies like OpenAI.
However, market analysts are still waiting for the next version of R1 – a potential R2 model, the launch of which many expected earlier this year. Local reports hint that the delay in launch is due to the founder of the company, Liang Wenfeng's persistent desire to refine the model. Moreover, he manages his profitable asset management business High-Flyer.
As previously reported by Cryptopolitan, DeepSeek postponed the launch of its AI-based R2 model due to ongoing technical issues with Huawei's Ascend processors. After the success of the R1 model in January, Chinese authorities recommended that DeepSeek use Huawei chips instead of American products from Nvidia. However, the company faced serious problems at the R2 model training stage. Sources familiar with the situation reported that DeepSeek had to use Nvidia chips for training and Huawei's Ascend processors only for data output. Industry insiders note that Chinese chips, including Huawei, often lag behind Nvidia in inter-chip communication quality, software support, and overall stability.
Huawei sent engineers to DeepSeek's offices to adapt the model. However, the startup could not successfully complete training on Ascend hardware even with off-site assistance. The originally scheduled launch of the R2 model in May was delayed due to these hardware issues.
Although some Chinese media speculate that the new model may be launched in the coming weeks, DeepSeek's founder Liang Wenfeng expressed internal dissatisfaction with the development process, urging the team to take the necessary time to develop a model that will maintain the company's competitive edge.
Meanwhile, industry giants like Alibaba and Tencent continue to actively release updates, with Alibaba's Qwen models attracting particularly many fans.
#MarketRebound , #Сryptomarketnews
If you cannot handle the large flow of news regarding cryptocurrencies and the world of finance, then perhaps you should not be subscribed to this group (!) ... 😉🙄
