Artificial Intelligence Insights

Research — JUNE 19, 2025 Generative AI Digest: A wave of notable AI model launches Chinese tech firms kicked off the month with notable AI model launches, including new releases from Alibaba Group Holding Ltd., Butterfly Effect and Hangzhou DeepSeek Artificial Intelligence Co. Ltd. These were followed by major announcements from Google LLC at its I/O conference and the long-anticipated debut of Claude 4 by Anthropic PBC. On the policy front, the US signaled a softer regulatory stance on semiconductors and AI, with a proposed decade-long pause on state legislation, and we saw a scaling back of proposed AI regulations in California. Meanwhile, the generative AI sector saw a flurry of acquisition activity. This month saw the long-awaited release of Claude 4 Opus and Sonnet from Anthropic, but the spotlight is on the growing momentum behind small language models and a novel diffusion-based approach for text generation. Microsoft Corp., Alibaba and ModelBest each introduced compact models with under 5 billion parameters, signaling a meaningful shift from when local deployment of language models required high-end hardware. Google made an array of announcements, including the unveiling of Gemini Diffusion. This experimental model applies diffusion techniques, which are traditionally used in image generation, to language tasks. By generating entire sections of text simultaneously and correcting errors mid-process, it breaks from the sequential process associated with conventional large language models (LLMs). The result is a dramatic speed boost. DeepMind Technologies Ltd. reports generation speeds of up to 1,479 tokens per second. For context, third-party benchmarks place Gemini 2.5 Flash at about 400 tokens per second and GPT-4o at roughly 150. Product releases and updates In April 2025, Alibaba released Qwen3, a family of open-source language models under the Apache 2.0 license, featuring six dense models ranging from 0.6 billion to 32 billion parameters and two Mixture-of-Experts models. These MoE models, Qwen3-235B-A22B and Qwen3-30B-A3B, have 22 billion and 3 billion active parameters, respectively. Qwen3 models support "thinking" and "nonthinking" modes, and delivered strong benchmark performance across coding, math and instruction-following tasks. There has been particular interest in the smaller models (0.6 billion, 1.7 billion and 4 billion) and the implications they may have for efficient local deployment. Google I/O 2025 delivered a wave of AI-related announcements, with the Gemini brand taking center stage throughout the keynote. Among the highlights was Veo 3, Google's latest video-generation model, which now supports not only high-quality video output but also synchronized audio. This marks a significant step forward in multimodal generation. Google also showcased how AI will be increasingly embedded across its productivity suite, with features like automated email cleanup drawing attention, along with browser agent Project Mariner. Another notable reveal was AI Mode, a new conversational search experience that reimagines how users interact with Google Search. DeepSeek Prover v2 is an open-source language model developed by Chinese AI company DeepSeek, building on the foundation of its popular V3 model. Prover v2 is designed specifically for generating mathematical proofs using Lean 4, a formal programming language tailored for theorem proving. The model approaches problems by decomposing complex proofs into smaller, manageable sub-goals, solving each step individually, and then assembling the results into a complete, verifiable proof. Anthropic released Claude Opus 4 and Claude Sonnet 4, showcasing notable upgrades in coding, reasoning and tool integration. Both models now support parallel tool execution and improved memory capabilities, enhancing their ability to handle complex multistep tasks. Claude Code has entered general availability, accompanied by new API features aimed at streamlining developer workflows. While the models perform well against benchmarks, the company has seen some criticism due to the pricing. Nutanix Inc. has officially launched the latest version of Nutanix Enterprise AI, emphasizing enhanced agentic capabilities enabled by the integration of NVIDIA Corp.'s new NIM and NeMo microservices on Kubernetes. Nutanix also supports NVIDIA's GPU Direct Storage, and its Nutanix Unified Storage product is certified on NVIDIA AI Data Platform, facilitating the development of agentic AI systems. Red Hat Inference Server, the newest addition to the Red Hat AI family, combines commercial hardening of vLLM with compression capabilities for LLMs. Red Hat also launched llm-d, a new open-source initiative intended to support distributed inferencing. It also announced its integration with the NVIDIA Enterprise AI Factory validated design. Microsoft announced a set of reasoning models as part of its Phi-4 small language model family. Phi-4-reasoning, Phi-4-reasoning-plus and Phi-4-mini-reasoning were revealed at the end of April. The base Phi-4 reasoning model weighs in at 14 billion parameters, with its "plus" variant designed to use more resources during inference (i.e., using more tokens to think through tasks). Mini reasoning is just 3.8 billion parameters, and Microsoft suggests it outcompetes open-source alternatives that have 7 billion or 8 billion parameters. Suno 4.5, the latest update to the AI-powered music creation tool, introduces broader genre support and significantly improved vocal quality. One of the standout upgrades is the ability to generate tracks up to eight minutes long. This is double the previous limit. The release also includes a new prompt enhancer, designed to refine user inputs and consequently improve the quality of the model's musical output. Chinese AI startup ModelBest has introduced MiniCPM-2B, a compact language model with just 2 billion parameters. Despite its size, the company claims it delivers performance on par with Falcon-40B, a model 20 times as large. If this is borne out, it highlights continued progress in efficient model design from Chinese technology companies. Ideogram 3.0, the latest version of the text-to-image-generation model, introduces improved photorealism and significantly better text rendering. A key new feature is Style References, which allows users to guide image generation using reference images. These styles can be saved and reused, allowing for more consistent and customizable outputs across projects. Hugging Face, Inc. has added computer use capabilities, smolagents, to its open-source agent library to help simplify agent development through higher-level abstractions. This feature enables models to perform human-like actions across digital interfaces, similar to tools from Anthropic, OpenAI LLC and Amazon.com Inc. As an open-source project, smolagents' expansion is expected to drive competition in the emerging "operator" space. Microsoft has also joined the trend, with Copilot Studio recently adding similar functionality. Manus AI, a general-purpose "computer use" AI agent product from startup Butterfly Effect, has entered general availability. The company showcased examples of the agent engaging in multi-step processes like building training materials for a topic, which it delivered as a website, or building a business case for additional software licenses. Funding and M&A Autonomous security vendor Torq Technologies Ltd. has announced its acquisition of Israeli startup Revrod Ltd., known for its work in multi-agent retrieval-augmented generation. The deal is positioned as a strategic technology addition, with Torq revealing that Revrod's capabilities are already integrated into Torq HyperSOC-2o, the latest iteration of its flagship platform. Salesforce Inc. has signed a definitive agreement to acquire Convergence Labs Ltd., an AI research lab based in the UK; the transaction is expected to close between April and June 2026. This acquisition aims to further enhance Salesforce's capabilities in agents, with Convergence Labs having focused on AI assistants delivering web tasks. LMArena, the organization behind the widely used Chatbot Arena platform for benchmarking AI models, has announced a $100 million series A funding round. Previously supported through grants and community donations, LMArena formally launched as a company in 2025. The new funding is expected to accelerate its expansion into domain-specific model evaluations. On May 16, 2025, Canadian AI model provider Cohere Inc. announced the completed acquisition of Ottogrid, a company specializing in AI-powered document analysis, data enrichment and research tools. Speech-to-text platform Rev.com, Inc. announced the acquisition of LegalDocumentGeneration, a company that did business as SmartDepo. The company provides technology for testimony analysis and suggests it will be able to increase attorney and court reporter productivity when paired with Rev's accurate transcription. AI21 Labs Ltd., an Israeli startup that recently expanded into AI orchestration, has announced a $300 million series D funding round. The round includes participation from a range of strategic and financial investors, including Alphabet Inc., NVIDIA and Intel Capital. This latest raise brings AI21 Labs' total funding to $636 million. Amazon Web Services Inc. and Saudi AI firm Humain have announced a strategic partnership to invest over $5 billion in developing an "AI Zone" in the Kingdom of Saudi Arabia, aligning with the Kingdom's Vision 2030 strategy. Separately, Humain and Advanced Micro Devices Inc. plan to invest up to $10 billion to deploy 500 MW of AI compute over the next five years. Cisco Systems Inc. has announced a collaboration with Humain focused on AI infrastructure. Telefonaktiebolaget LM Ericsson (publ) has joined a consortium with AstraZeneca PLC, Saab AB (publ), SEB and Wallenberg Investments to establish an AI factory and deploy an enterprise AI supercomputer in Sweden. Politics and regulations The current US administration has officially rescinded the Biden-era "AI diffusion rule," which had restricted exports of advanced AI technologies, particularly chips. A Commerce Department spokesperson said the rule will be replaced with a streamlined framework aimed at boosting US innovation and maintaining AI leadership. The original policy faced strong opposition from major tech firms like Microsoft and NVIDIA, which argued that it limited global competitiveness. Meanwhile, the One Big Beautiful Bill Act of 2025, which includes a controversial 10-year ban on state-level AI regulation, has passed the House but is expected to face challenges in the Senate. Critics warn that, in the absence of federal oversight, the moratorium could leave the US AI ecosystem with minimal safeguards. Japan's lower house has passed a bill aimed at promoting AI development while addressing associated risks. Expected to be enacted by the end of the current parliamentary session in June, the legislation establishes a prime minister-led task force to draft a national AI strategy. While it includes provisions for addressing AI misuse, there are no explicit penalties, although authorities may publicly disclose malicious actors. Following enactment, the government will issue guidelines for AI developers aligned with international norms, including the use of digital watermarks. Under pressure from California Governor Gavin Newsom and business groups, the California Privacy Protection Agency has significantly scaled back its proposed AI regulations. In a unanimous vote, the agency removed key provisions, including those targeting behavioral advertising, and narrowed the scope from broadly regulating "artificial intelligence" to focusing on "automated decision-making." This shift is expected to exempt many applications from oversight. Gain access to our full news .& research coverage and the industry-specific data that informs our insights This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global. S&P Global Market Intelligence 451 Research is a technology research group within S&P Global Market Intelligence. For more about the group, please refer to the 451 Research overview and contact page.