I ditched LM studio for an open-source alternative — and my local model is doing things it couldn't before

At a glance:

• Switched from LM Studio to the open‑source Jan app for a GUI that supports audio and advanced MCP browser tools out of the box.
• Jan can run both local GGUF models (e.g., Gemma 4 E4B) and cloud APIs such as Claude or ChatGPT from a single interface.
• The new workflow eliminates manual JSON edits, lets users import existing llama.cpp models, and offers deeper sampler controls like DRY and Mirostat.

Why the switch

The author’s journey began with LM Studio because its one‑click installer and clean graphical interface made self‑hosting LLMs feel approachable. Over time, however, the proprietary tool proved too conservative: newer features such as multi‑call‑point (MCP) support, audio handling, and the latest llama.cpp samplers were either missing or required cumbersome manual configuration. The need to repeatedly wire up browser automation and edit JSON files became a productivity bottleneck, prompting a search for a more modern, open‑source alternative.

Jan’s feature set

Jan positions itself as a hybrid between LM Studio’s user‑friendly GUI and llama.cpp’s raw power. It ships as a desktop app that is fully open‑source, with every chat, configuration, and model stored as plain files on the user’s disk. The UI mimics ChatGPT for familiarity, while the underlying philosophy is “local‑first, file‑over‑app,” meaning users retain full control over their data. Importantly, Jan includes built‑in audio support, allowing models with speech capabilities to be used without external plugins—a capability LM Studio lacks.

Bringing my own model

To test Jan, the author imported a Gemma 4 E4B GGUF model from Hugging Face. The model, chosen for its speed and multimodal abilities on an 8 GB VRAM system, was moved into a clean folder and linked directly via Jan’s “Llama.cpp Import” feature. Unlike LM Studio, Jan does not copy the model files; it references them in place, so the user must keep the folder intact. Once linked, the model could accept audio, image, and document inputs, showcasing Jan’s broader modality support.

MCP browser setup that actually makes sense

Jan’s Browser MCP server arrives pre‑configured and runs through an npx bridge to a Chrome extension. Activation is as simple as installing the extension and toggling the server in Jan’s settings—no JSON editing, no manual API‑key registration. The author compared this to LM Studio’s Brave Search MCP, which required hand‑editing JSON and debugging restarts. Jan’s MCP works in any Chromium‑based browser; the author locked down permissions to “on click” and disabled it in private/Tor windows to mitigate prompt‑injection risks. The result was a smooth, buttery browsing experience where the model could fetch live data instantly.

Remote providers and unified API keys

Although marketed as a local‑first tool, Jan includes roughly a dozen remote providers out of the box. Users can paste API keys for services like Claude or ChatGPT and interact with those clouds alongside locally hosted models. A Hugging Face Router integration also offers pay‑as‑you‑go access to hosted models such as Kimi‑K2 and DeepSeek‑R1 without needing separate accounts for each provider. This dual capability streamlines workflows for developers who toggle between on‑device inference and cloud‑based power.

Deeper controls and future outlook

Jan exposes advanced llama.cpp samplers—DRY, Mirostat, dynamic temperature—and lower‑level knobs such as tensor buffer overrides and mixture‑of‑experts (MoE) placement directly within the GUI’s model settings pop‑over. These controls were previously hidden behind command‑line flags or unavailable in LM Studio. By bundling them into an accessible interface, Jan lowers the barrier for power users while preserving the openness of the underlying codebase. The author concludes that Jan represents the “perfect middle ground” for anyone who values both ease of use and the transparency of open‑source AI tooling.