LM Studio

A desktop workbench for running open-source models locally, with the new Bionic agent for coding, document processing, and local voice input

LM Studio is a desktop workbench that runs open-source LLMs right on your own computer: a graphical interface for browsing, downloading, and loading open-weight models such as Qwen, Llama, GLM, and Kimi; full coverage across macOS/Windows/Linux; and inference that never leaves your machine. Bionic, released on July 16, upgrades it from a “model runner” into an “AI agent” — it can inspect local codebases, explain unfamiliar code, edit files directly with inline diffs, and also handles documents, spreadsheets, and slides.

It fits two kinds of users: developers and researchers who value privacy and don’t want code or documents leaving their machines; and anyone who wants to keep AI spending under control — light tasks run on small local models at zero cost, while heavy lifting goes through Secure Cloud to open-source flagships like GLM 5.2 and Kimi K2.7 Code on a pay-as-you-go basis. The official promise: zero data retention, and no training on user data. The voice keyboard launches with Mistral’s Voxtral for local real-time transcription, letting you dictate by voice in any app.

Where Ollama takes the CLI-first, hacker-oriented route, LM Studio goes the polished, “out-of-the-box” product path; and compared with the ChatGPT/Claude desktop apps, it bets entirely on the open-weight ecosystem. On its release day, Bionic hit the Hacker News front page (300+ points) and is being tipped as a candidate for the “Cursor moment of open-source models.”

Note: Running large models locally is memory-intensive (16GB+ unified memory or an equivalent GPU recommended); Bionic’s cloud calls are billed by usage, so keep an eye on your model and compute environment choices in Settings; and commercial licensing terms for some flagship open-source models should be confirmed on their download pages.

Source: Agentopia · Full-site machine-readable catalog at /llms.txt