Unsloth Studio just shipped a meaningful update across four fronts: MCP, Projects, Canvas, and a redesigned chat UI. Gemma 4 12B support is also live. If you build with local models, each of these touches your daily workflow.
The headline addition is MCP support. One click in the composer gives your model live tools without needing to supply API keys for the built-ins. Three presets ship out of the box: Context7 for current library docs and code, Exa for live web search, and Hugging Face for searching models, datasets, and papers. You can also add your own remote servers (OAuth or headers) or local servers (stdio), toggled per chat. That last detail matters. You are not committing globally. Each chat controls its own tool surface.
Projects let you group related chats into a shared workspace. Create a project from the sidebar, then attach new or existing chats to it. Simple, but it solves the context sprawl problem that hits anyone running multiple experiments or client threads in parallel.
Canvas is experimental but worth watching. It opens generated HTML inside a dedicated panel in Unsloth Studio, renders interactive browser-based visualizations, supports CDN-loaded packages, and lets you flip between the rendered preview and raw source. For anyone building or demoing data tools, that is a meaningful shortcut.
The new chat UI consolidates Projects, Canvas, MCP, and Compare into a single + menu. Search and Code are now single-click. Less navigation overhead during active work.
On the model side, Google's Gemma 4 12B is now supported. It runs locally on 8GB RAM and handles image input, audio input, and up to 256K context. You can run it or fine-tune it directly inside Unsloth Studio. Find the GGUF builds on Hugging Face and the setup guide in the Unsloth docs.
A few infrastructure notes: CUDA 13.3 llama.cpp binaries now work on Windows and fix the gibberish-output bug present in CUDA 13.2. The default still pins to CUDA 13.1 for now. On CUDA 13.2, 13.1, and below, Windows falls back to CUDA 12.4.
One critical install note: update via the install script, not via unsloth studio update. The update command will not pull the latest packaging changes. On macOS, Linux, or WSL, run curl -fsSL https://unsloth.ai/install.sh | sh. On Windows, run irm https://unsloth.ai/install.ps1 | iex. If Gemma 4 12B is not loading correctly, a fresh install via the script is the fix.
The concrete action today: if you have been manually wiring tool calls or juggling context across unrelated chats, update Unsloth Studio using the script above, enable MCP presets for your next session, and try grouping your active workstreams into Projects.