This week, DeepSeek released an upgraded version of its R1 model on HuggingFace, reigniting the open-source versus closed-source discussion. The latest iteration, DeepSeek-R1-0528, features 685 billion parameters, up from 671 billion in January. In contrast to the closed-source models from OpenAI and Google, the model weights of DeepSeek are publicly available. Benchmark results suggest that the R1-0528 update has improved reasoning and inference capabilities, reducing the disparity with OpenAI’s o3 and Google’s Gemini 2.5 Pro.

DeepSeek also introduced a distilled variant of R1-0528 using Alibaba’s Qwen3 8B model, providing a lighter option that demands less computing resources. The DeepSeek-R1-0528-Qwen3-8B outperforms Google’s Gemini-2.5-Flash-Thinking-0520 and OpenAI’s o3-mini in various benchmarks. Notably, the distilled model from DeepSeek can operate on a single GPU, as noted by TechCrunch.

In conclusion, the Chinese competitor is making strides against its U.S. rivals with a strategy focused on open weights that is more affordable and accessible. DeepSeek illustrates that AI models may not require the extensive computing power currently employed by OpenAI, Google, and others. However, DeepSeek’s models come with their own set of challenges. An AI developer quoted by TechCrunch stated that the latest DeepSeek update is even more restricted than its predecessor regarding critiques of the Chinese government.

In other AI developments, following last week’s surge of announcements from Google, Anthropic, and Microsoft, this week was less eventful in terms of product updates and feature rolls. This allowed significant focus on DeepSeek’s R1 update. Furthermore, Anthropic launched voice mode, AI influencers became more popular, Anthropic’s CEO cautioned about potential mass layoffs, and an AI-generated kangaroo drew attention.

Google’s Veo 3 has gone viral, as users across social media express their enthusiasm for the new AI video model. The impressive outcomes have already resulted in short films created entirely with Veo 3, merely 11 days post-release.

In additional AI product news, Anthropic rolled out voice mode for Claude, now in beta on mobile for the Claude app, allowing users to summarize calendars or read documents aloud. OpenAI is investigating the potential to sign in to third-party applications using ChatGPT, which may connect personalized ChatGPT settings to other apps. Opera unveiled a new agentic AI browser named Neon, featuring a chatbot interface and capabilities for filling out web forms.

The browser frontier for agentic AI is expanding, with Perplexity developing a similar tool named Comet, and The Browser Company shifting focus from its Arc browser to a more AI-driven browser called Dia. Meanwhile, Google may have to divest Chrome, with OpenAI expressing interest in acquiring it.

Anthropic CEO Dario Amodei has cautioned that AI could displace half of all entry-level white-collar positions. A report from VC firm SignalFire indicated a decline in hiring for entry-level roles, partly due to AI automating less-complex tasks.

In AI culture, Google claims its AI overviews engage 1.5 billion people monthly, though they still encounter difficulties with basic functions. A suggested 10-year prohibition on state regulation of AI is unpopular, with a poll indicating 57 percent opposition. The New York Times has established a licensing agreement with Amazon for AI model content, despite suing OpenAI over copyright violations.

A viral video featuring an emotional support kangaroo was generated by AI, and AI is also being utilized to depict public figures as infants. On a positive note, HuggingFace has launched a humanoid robot named HopeJr, while the Colossal Foundation has developed an AI algorithm to detect the calls of the nearly extinct tooth-billed pigeon.

For the latest updates in AI, check back next week for another recap, and follow @cecily_mauran and @mashable for more information.

Disclosure: Ziff Davis, Mashable’s parent company, initiated legal action against OpenAI in April, claiming copyright infringements associated with the training and operation of its AI systems.