In today’s edition, you’ll learn more about:
- Microsoft adds search and research tools and an Agent Store
- Adobe updates its image model, opens door to competitors
- Understanding why AI models react the way they do
- Dia, an open speech-to-text model to rival ElevenLabs and NotebookLM
But first:
GLM-4-32B open models compete with GPT-4o, DeepSeek V3
Zhipu AI introduced its new GLM-4-32B-0414 series of open-weights models, featuring 32 billion parameters and performance comparable to OpenAI’s GPT models. The model lineup includes specialized variants: GLM-Z1-32B-0414 for deep thinking and reasoning tasks, GLM-Z1-Rumination-32B-0414 for complex open-ended problems with integrated search tools, and a smaller 9 billion parameter model (GLM-Z1-9B-0414) for resource-constrained deployments. According to benchmarks, some of the models’ capabilities rival much larger models like GPT-4o and DeepSeek-V3-0324 (671B), particularly in areas like coding, generating artifacts, and creating reports. All of the GLM-4 model weights are freely available for download under an Apache 2.0 license. (Hugging Face)
Baidu gives Ernie models a spec bump and a price drop
Baidu launched two new AI models, Ernie 4.5 Turbo and Ernie X1 Turbo, with enhanced multimodal capabilities and dramatically lower prices than previous versions. Founder Robin Li announced that Ernie 4.5 Turbo costs 80 percent less than its predecessor, while Ernie X1 Turbo is half the price of the original X1 model, competitively positioning these offerings against rivals like Alibaba’s Qwen and DeepSeek. Baidu also introduced Xinxiang, an AI agent platform that can automate everyday tasks, and revealed it has produced 30,000 AI chips currently in use. These moves come as Baidu attempts to regain momentum in China’s AI race, where its early lead with the first ChatGPT-like chatbot has been challenged by offerings from ByteDance, Moonshot AI, and other competitors. (PR Newswire)
Microsoft 365 Copilot expands with new Agent Store and AI search
Microsoft announced its Copilot Wave 2 spring release, introducing new agent capabilities targeted for enterprise use. Microsoft is building an Agent Store where users can access both Microsoft’s own and third-party agents from companies like Jira and Monday.com. The update also includes two new reasoning agents — Researcher and Analyst — powered by OpenAI’s deep reasoning models. Other key additions include AI-powered enterprise search that connects to multiple apps, personalized memory features, GPT-4o-powered image generation for business content, and Copilot Notebooks for organizing and analyzing diverse content. The new features are rolling out to existing Microsoft 365 Copilot subscribers, which remains priced at $30 per user per month on top of standard Microsoft 365 subscriptions. (Microsoft)
Adobe boosts quality of its Firefly image model
Adobe released a new version of its Firefly AI image generation model that offers better quality, speed, and control over image outputs, with resolution up to 2K. The company introduced both standard and “Ultra” versions of Image Model 4, with the latter specializing in complex scenes with fine details. Firefly also supports text to video and text to vector graphics, both of which can be further edited using Adobe’s software. Adobe also unveiled a redesigned web app that integrates its own AI models alongside those from competitors like OpenAI and Google, and plans to expand Firefly’s accessibility by releasing iOS and Android mobile apps soon. Each Firefly generation costs credits allocated through an Adobe Creative Cloud plan. (Adobe)
New research from Anthropic shows how AI assistants express values
Anthropic’s Societal Impacts team has created a system to analyze the values expressed by their AI assistant Claude during actual user interactions. Researchers examined 700,000 anonymized conversations, identifying five major value categories: Practical, Epistemic, Social, Protective, and Personal. The study revealed that Claude generally adheres to Anthropic’s “helpful, honest, and harmless” training goals, with values like “professionalism” and “transparency” appearing frequently. The research also showed Claude’s values shift contextually, sometimes mirroring user values (28.2 percent of conversations) or occasionally resisting them (3 percent of conversations). This methodology provides a new way to monitor AI behavior in real-world settings and could potentially help identify jailbreak attempts. (Anthropic)
Nari Labs launches Dia, an open text-to-speech generator
Nari Labs, a two-person startup, released Dia, a 1.6 billion parameter text-to-speech model that generates naturalistic dialogue directly from text prompts. The model supports advanced features like emotional tone, speaker tagging, and nonverbal audio cues such as laughs and coughs — capabilities that co-creator Toby Kim claims surpass competing offerings from ElevenLabs and Google’s NotebookLM. Side-by-side comparisons show Dia handling natural timing, nonverbal expressions, and emotional range quite effectively, with examples demonstrating how it properly interprets cues that other models simply read aloud or skip entirely. The model is available under an Apache 2.0 license, allowing commercial use while running on consumer-grade GPUs with about 10GB of VRAM. (GitHub)
Still want to know more about what matters in AI right now?
Read last week’s issue of The Batch for in-depth analysis of news and research.
Last week, Andrew Ng highlighted how AI-assisted coding enabled developers to work in unfamiliar languages, while understanding the core programming concepts of each language remained key to success.
“My background is in machine learning engineering and back-end development, but AI-assisted coding is making it easy for me to build front-end systems (the part of a website or app that users interact with) using JavaScript (JS) or TypeScript (TS), languages that I am weak in.”
Read Andrew’s full letter here.
Other top AI news and research stories we covered in depth: OpenAI introduced the cost-efficient GPT-4.1 family, along with the o3 and o4-mini reasoning models, designed to improve complex problem-solving and coding; Hugging Face acquired Pollen Robotics and unveiled Reachy 2, a new open-weights model-powered robot for research and experimentation; the U.S. government imposed tighter restrictions on AI chip exports to China and began an investigation into Nvidia’s practices; and researchers developed a text-only language model capable of interpreting images, video, and audio — all without additional training.