Gemma 4 and Cursor 3 Drop Today: A Dev Roundup

Google dropped Gemma 4 today, and Cursor shipped version 3 with a complete interface overhaul. Both announcements landed within hours of each other, and they point in the same direction: the AI tooling stack is shifting hard toward agent-centric workflows and open, run-anywhere models.

TL;DR: Gemma 4 brings Google's best open-weight models yet under Apache 2.0, with vision, function-calling, and sizes from 2B to 31B. Cursor 3 reimagines the IDE around parallel agents that work across repos, locally, or in the cloud.

Gemma 4: Open Models That Actually Compete

Google DeepMind released Gemma 4 today, and the numbers are striking. The 31B variant ranks #3 on the Arena AI text leaderboard. The 26B Mixture-of-Experts model sits at #6. Both are beating models twenty times their size.

Four variants are available now:

• E2B and E4B (Effective parameters) — built for on-device and edge deployment, 128K context
• 26B MoE — mixture-of-experts architecture, 17.99GB in GGUF format
• 31B Dense — flagship dense model, 19.89GB, fits on a single 80GB H100

All models ship under Apache 2.0, which means no usage restrictions, no royalty headaches. Simon Willison already ran pelican SVG generation tests across the lineup and documented the quality progression from 2B up to 26B MoE.

The smaller models use Per-Layer Embeddings (PLE), a technique that gives each decoder layer its own embedding table for every token. This keeps effective parameter counts low while maintaining capability. Think of it as a lookup shortcut that trades memory for compute efficiency.

Key capabilities across the family:

• Vision and audio input on all models
• Function-calling and structured JSON output for agent workflows
• 256K context on the larger variants
• High-quality offline code generation

You can grab the GGUF files from Hugging Face and run them locally with LM Studio or Ollama. API access to the 31B model is available through Google AI Studio.

Cursor 3: The IDE Becomes an Agent Dashboard

Cursor shipped version 3 today with a fundamental rethinking of how an AI IDE should work. The old VS Code fork is gone. In its place is a new interface built from scratch, centered entirely around agents.

The headline feature is the Agents Window. It lets you run multiple agents in parallel across different repositories and environments. Local, worktree, cloud, SSH — all visible in one sidebar.

Cloud agents now produce demos and screenshots of their work so you can verify outputs without digging through logs. When you want to dig deeper, the IDE is still there. You can switch back anytime or run both side by side.

The handoff between local and cloud is where this gets interesting. Move an agent from cloud to local when you want to test edits on your own hardware. Move from local to cloud when you need a task to keep running while your laptop is closed. Composer 2, Cursor's frontier coding model, handles the iteration loop.

The new diffs view simplifies PR workflows. Stage, commit, and manage pull requests without bouncing to GitHub. It's a small change, but it keeps you in the flow.

Self-hosted cloud agents also shipped recently, letting enterprises keep code and tool execution inside their own network. No data leaves your infrastructure. This matters for organizations with strict data residency requirements.

What This Means for Developer Workflows

Gemma 4 and Cursor 3 are solving different problems, but they reinforce each other. Better open models mean more capable local agents. Better agent UX means those models actually get used in production workflows.

The pattern is clear: the action is moving from chat interfaces to persistent, tool-using agents that run across your infrastructure. Whether you're fine-tuning a 4B model for your specific use case or orchestrating a fleet of coding agents across repos, the tooling is finally catching up.

You can run Gemma 4 locally right now. Cursor 3 is available for download. The model weights are on Hugging Face. No waiting required.

Sources: Google DeepMind Gemma 4, Simon Willison, Cursor 3 Announcement, Hacker News