The big theme this month was turning LMPlayground from a chat app into an agentic, multimodal one. 1.7.0 gave capable models tools — web search, page fetch, and a JavaScript sandbox they can call mid-reply. 1.8.0 then added on-device vision: attach a photo (or snap one) and chat about it with Gemma, Qwen, or Ministral vision models, all running locally.
Both Spark projects shipped 1.2.2, whose headline is Windows support — the CLI, the skills, and the Claude connector now run on Windows alongside macOS.
LMPlayground
- 1.7.0 — Tools — capable models can now search the web, fetch a page, and run JavaScript during a reply, toggled per model in Settings → Tools.
- 1.7.1 — Added Gemma 4 12B (Google) and Granite 4.1 3B & 8B (IBM) to the model catalog.
- 1.8.0 — Vision — send or snap a photo and ask your model what it sees, with Gemma, Qwen, and Ministral vision models, running fully on-device. Enabled the Vulkan GPU backend.
spark-cli-skills
- 1.2.2 — Windows support for the Spark CLI, in addition to macOS. Added
recipe-unsubscribe-audit— a full unsubscribe audit that scans inbox, archive, and GateKeeper for newsletter senders, classifies by volume and engagement, flags phishing and duplicates, then unsubscribes (contributed by @quser).
spark-claude-extension
- 1.2.2 — Windows support — the Claude connector now runs on Windows alongside macOS.