Google DeepMind has released Gemma 4 12B, featuring an innovative encoder-free unified architecture that directly integrates visual and audio inputs into the language model backbone, eliminating the need for traditional multimodal encoders and reducing latency and memory usage. The model achieves performance close to a 26B MoE model while consuming less than half the memory, enabling local execution on consumer devices with only 16GB of VRAM. It is the first mid-sized Gemma model to natively support audio input and includes multi-token prediction technology to reduce inference latency. Released under the Apache 2.0 license, the Gemma series has surpassed 150 million cumulative downloads. Google also launched an official Skills library for agent development and a macOS desktop app enabling fully localized voice and visual interaction.

Chinese AI startup DeepSeek is nearing completion of its first external funding round, raising approximately $7 billion (50 billion RMB), with a valuation between $52 billion and $59 billion. Founder Liang Wengao personally contributed 20 billion RMB, taking the lead stake, while Tencent and CATL plan to invest around 10 billion and 5 billion RMB respectively, becoming the largest external investors. NetEase and JD.com also participated. This round marks DeepSeek’s transition from a founder-funded research project to a commercial entity accountable to investors. Despite a valuation far below OpenAI or Anthropic, DeepSeek has gained international recognition for training high-performance open-source models at low cost. The core challenge now lies in maintaining its openness and efficiency after securing substantial external capital.

Ideogram AI has released Ideogram 4.0, a 9.3B-parameter model based on a full single-stream Diffusion Transformer architecture, using Qwen3-VL-8B-Instruct as the text encoder. The model supports native 2K resolution output, multilingual text rendering, bounding box layout control, and color palette specification. It ranks first among open-source models on Design Arena and LMArena, trailing only closed-source models like GPT and Gemini. In professional designer evaluations, it led FLUX.2 and Gemini Flash with a 47.9% preference rate. Model weights are released in nf4 and fp8 quantized formats on Hugging Face and GitHub under a non-commercial license. The API offers three quality tiers: Turbo at $0.03 per image and Default at $0.06 per image.

AI music generation company Suno has completed a $400 million Series D funding round at a $5.4 billion valuation—more than doubling its $2.45 billion valuation from seven months prior—despite facing multiple copyright lawsuits. Bond Capital led the round, with participation from existing investors. Suno acknowledges its models were trained on copyrighted songs and defends this under fair use, but Universal Music Group, Sony Music, and others continue legal action; plaintiffs have recently alleged unauthorized use of over 61,000 songs. Warner Music reached a settlement and signed a licensing agreement with Suno in November 2025. Suno users generate over 7 million songs daily, and the app remains a top performer in the App Store's music category.

Meta has announced the global rollout of Meta Business Agent, an AI customer service agent for WhatsApp Business that can automatically answer inquiries, recommend products, schedule appointments, qualify sales leads, and escalate to human agents when needed. The feature is now also integrated into Instagram Direct Messages. Meta is testing AI-generated daily chat summaries and business insights, with plans to expand into market research and competitor analysis. The company is also building a platform for large enterprises to connect systems like Shopify and Zendesk to create custom AI agents. The service will be part of the WhatsApp Business Premium subscription, with enterprise pricing based on token usage. Meta's stock rose nearly 3% on the announcement day.

NVIDIA has acquired four-year-old AI startup Kumo AI, which specializes in developing foundational models for precise commercial forecasting. Kumo AI's co-founders include CEO Vanja Josifovski, Engineering Lead Hema Raghavan, and Chief Scientist Jure Leskovec. The acquisition aims to enhance NVIDIA’s technical capabilities in AI and continues its recent strategy of active M&A in the artificial intelligence space. Financial terms were not disclosed.

The European Council has reached a provisional agreement on the 'Digital Package,' making key amendments to the EU AI Act. Compliance deadlines for high-risk AI systems are postponed to December 2, 2027; those embedded in products must comply by August 2, 2028; and national AI regulatory sandboxes must be established by August 2027. Meanwhile, transparency requirements for generative AI content are accelerated, requiring providers to implement measures by December 2, 2026. The Act introduces new prohibitions, explicitly banning the use of AI to generate non-consensual intimate images and child sexual abuse material. Consulting firm ADC warns that delayed enforcement could increase corporate compliance costs and leave the public exposed to poorly regulated AI systems for longer.

OpenAI has publicly shared its detailed research roadmap following organizational restructuring, aiming to achieve a fully autonomous AI researcher capable of independently completing large-scale research projects by March 2028. Chief Scientist Jakub Pachocki stated that deep learning may be less than a decade away from superintelligence and introduced a new Chain-of-Thought fidelity technique to improve value alignment. Sam Altman acknowledged issues with GPT-4o’s forced routing mechanism in sensitive conversations. Under the new structure, the nonprofit OpenAI Foundation will govern the public benefit corporation, allocating $25 billion toward AI-driven healthcare research and AI resilience initiatives. OpenAI also disclosed commitments exceeding 30GW of infrastructure investment, representing total financial obligations of approximately $1.4 trillion. On the same day, OpenAI announced a major upgrade to its GPT-Rosalind life sciences model, enhancing drug discovery and experimental workflow capabilities.

Researchers at the University of Toronto have demonstrated a prototype worm virus powered by publicly available AI models, capable of exploiting known computer vulnerabilities to autonomously spread across Linux, Windows, and IoT devices without human intervention. The worm can steal passwords from infected devices, discover new vulnerabilities, and leverage device compute power to plan subsequent attacks, significantly lowering attack costs. While the prototype currently exploits only known vulnerabilities, researchers warn malicious actors could adapt it to autonomously discover zero-day exploits. They emphasize that in an interconnected world, no system is immune to such threats and urge rapid action from all stakeholders.

Anthropic has selected Morgan Stanley and Goldman Sachs as joint lead underwriters for its IPO, with JPMorgan also participating. The company has secretly filed a registration draft with the SEC and may launch its public offering in October, targeting a valuation of approximately $1 trillion and aiming to raise up to $60 billion. Prior to this, Anthropic raised $65 billion in its Series H round, achieving a post-money valuation of $965 billion. SpaceX’s S-1 filing revealed it supplies Anthropic with about 325,000 NVIDIA chips monthly at a cost of roughly $1.25 billion, under a contract running through May 2029. Anthropic’s annualized revenue has reached $30 billion, marking an 80x growth. Meanwhile, 55% of enterprises cite reliability and hallucination management as primary barriers to adopting generative AI.