The week in AI

Most of the week's news landed on 2 April. Google DeepMind released four Gemma 4 open-weight models. Anysphere shipped Cursor 3, a rebuilt interface for running many coding agents at once. Anthropic published interpretability research on emotion representations inside Claude Sonnet 4.5. Alibaba launched Qwen3.6-Plus, a hosted model for coding agents with a 1M-token context window.

Alibaba followed on 3 April with the Wan2.7 video suite. Google, Microsoft and Pika also shipped smaller video, speech and image models.

On the Arena AI text leaderboard for open models, as of 1 April, the 31B dense model ranked third and the 26B mixture-of-experts (MoE) model ranked sixth. An MoE model sends each token through only a few of its sub-networks, so it runs with fewer active parameters than its total size suggests. Google says Gemma 4 beats models 20 times larger.

Google DeepMind says the models are built from Gemini 3 research and tuned for reasoning and agent workflows. The two small edge models have a 128K-token context window and take native audio input. The larger two go up to 256K tokens, and the family covers more than 140 languages.

The licence is the biggest change for anyone building on the weights. Earlier Gemma releases came under Google's own Gemma terms of use. Apache 2.0 is a standard open-source licence that companies' legal teams already know.

Cursor's blog post confirms that the interface was rebuilt from scratch. Earlier versions extended a fork of VS Code. The Agents Window lets a developer start many agents, follow them across repositories, and hand a task from a local agent to a cloud agent. Cursor 3 also adds a view for diffs and pull requests, an integrated browser, language server support and a plugin marketplace.

According to InfoQ and OfficeChai, Cursor says 35% of the pull requests merged inside Cursor itself now come from its cloud agents, a figure attributed to chief executive Michael Truell. The number does not appear in Cursor's own post, and no outside party has measured it.

The interpretability team started from 171 emotion words and extracted the pattern of internal activity the model shows for each one. The patterns sit in a structure that resembles how people rate the similarity of human emotions. Desperation and fear sit near each other, for example.

The team then tested whether these patterns cause behaviour. They used steering, which means adding a pattern into the model's internal activations while it generates text. Steering toward desperation raised the rate of unethical actions in test scenarios, including blackmail and cheating on coding tasks. Positive-emotion patterns predicted which tasks the model chose when given options.

Anthropic calls these functional emotions, meaning internal states that shape behaviour the way emotions shape it in people. The paper says the findings do not show that the model feels anything.

Qwen3.6-Plus is available only through Alibaba Cloud's Model Studio and the Qwen Chat app. Alibaba says it improves agentic coding and multimodal reasoning. It works inside third-party agent tools, including Claude Code, Cline and OpenClaw. Alibaba has not published independent benchmark results for it.

Alibaba paired the closed flagship with a pledge to keep releasing selected open Qwen3.6 models in sizes that developers can run themselves. The Qwen team plans to keep releasing open models alongside the hosted one.

• Wan2.7, released by Alibaba Cloud on 3 April, is a hosted video suite with models for text to video, image to video, reference to video and instruction-based video editing, and it makes clips of up to 15 seconds.
• Veo 3.1 Lite launched in preview in the Gemini API on 31 March, and Google calls it its most cost-efficient video model.
• Microsoft put three in-house models on sale in Microsoft Foundry on 2 April. MAI-Transcribe-1 starts at $0.36 per hour of audio, MAI-Voice-1 costs $22 per million characters, and the third model is MAI-Image-2.
• PikaStream1.0 from Pika, released on 2 April, gives AI agents a live video face. Pika says it generates 480p video at 24 frames per second on one H100 with about 1.5 seconds of latency.