Today we are thrilled to share LM Studio 0.4.0, the next generation of LM Studio.
This release introduces parallel requests with continuous batching for high throughput serving, all-new non-GUI deployment option, new stateful REST API, and a refreshed user interface.
LM Studio 0.4.0 highlights include:
- Deploy LM Studio's core on cloud servers, in CI, or anywhere without GUI.
- Parallel requests to the same model with continuous batching (instead of queueing).
- New stateful REST API endpoint:
/v1/chatthat allows using local MCPs. - Refreshed application UI with chat export, split view, developer mode, and in-app docs.
Read on for more details!
Today we're introducing llmster: it's the core of the LM Studio desktop app, but packaged to be server-native, without reliance on the GUI. We've rearchitected our software to separate the GUI from the core functionality, allowing llmster to run as a standalone daemon.
This means llmster can be run completely independently of the app and deployed anywhere: Linux boxes, cloud servers, your GPU rig, or even Google Colabs. It can of course still be run on your local machine without the GUI, for those who prefer terminal-based workflows.
How to install llmster
Linux / Mac
curl -fsSL https://lmstudio.ai/install.sh | bash
Windows
irm https://lmstudio.ai/install.ps1 | iex
Using llmster
- Start the daemon:
lms daemon up - Download a model:
lms get <model> - Start the local server:
lms server start - Open an interactive session:
lms chat - Update your runtime:
lms runtime update llama.cpp(andlms runtime update mlxon macOS)
Alongside LM Studio 0.4.0, our llama.cpp engine is graduating to version 2.0.0. With it we're introducing support for concurrent inference requests to the same model.
Run parallel requests in the app with Split View
Max Concurrent Predictions and Unified KV Cache
You will find 2 new load options in the model loader dialog:
-
Max Concurrent Predictions: sets the maximum number of concurrent requests that can be processed by the model. Requests beyond this limit will be queued.
-
Unified KV Cache: when enabled, preallocated resources will not be hard-partitioned per concurrent request, allowing varying request sizes per request. This is enabled by default.
Parallel requests work thanks to llama.cpp's open-source continuous batching implementation, adopted in LM Studio's llm-engine. This capability has not yet made it into our MLX engine, but it is actively in the works and will land soon.
We have refreshed LM Studio's user interface from the ground up for a more consistent and pleasant experience.
Export chats to PDF, markdown, or text
You can now export your chats to PDF, markdown, or plain text. Click the ••• menu on a chat and head to "Export" for all available options.
Split View
You can now open multiple chat sessions side by side using Split View. Click the new Split View icon in the top right corner of the chat window to open a new chat pane.
Developer Mode
Developer Mode is a new setting that exposes advanced options in the app. You can enable it from Settings > Developer. Once enabled, it'll reveal all advanced options across the app, including in the model loader dialog and sidebars.
In-app docs
Head over to the Developer tab to see the new in-app documentation. It covers the new REST API, CLI commands, and advanced configuration options.
Introducing lms chat
With LM Studio 0.4.0, we're introducing a brand-new CLI experience centered around the lms chat command. This command opens an interactive chat session directly in your terminal, allowing you to chat with your models and download new ones.
Run lms chat --help to see all available options.
/v1/chat endpoint
/v1/chat is a new first-party REST endpoint for chatting with local models from your apps.
Unlike typical "stateless" chat APIs, /v1/chat is stateful: you can start a conversation, get back a response_id, and then continue it by passing previous_response_id on your next request. This keeps requests small and makes it easy to build multi-step workflows on top of LM Studio.
Responses also include detailed stats (tokens in/out, speed, time to first token), so you can track performance and tune load/inference settings.
And when you need tools, /v1/chat can also enable your locally configured MCPs - gated by permission keys.
Permission keys
To allow you to control which client accesses your LM Studio server, we've introduced permission keys. You can generate and manage permission keys from the Settings > Server tab in the app.
Please let us know how you like it! We'd love to hear your feedback.
Special thanks to the 0.4.0 beta group. Your feedback and bug reports have been invaluable <3.
Below is the full list of release notes items.
### LM Studio 0.4.0 - Release Notes Welcome to LM Studio 0.4.0 👾! - We're excited to introduce the next generation of LM Studio. - New features include: - `llmster`: the LM Studio Daemon for headless deployments w/o GUI on servers or cloud instances - Parallel inference requests (instead of queued) for high throughput use cases - New stateful REST API with local MCP server support - `POST /v1/chat` - A completely revamped UI experience ✨ **Build 17** - MCPs will now only be loaded when needed, instead of at app startup - Fixed a bug where some fields in app settings could get reset after update **Build 16** - New icons and placements for Discover, My Models buttons - Fixed a bug where generators wouldn't show in the top bar model picker when selected - Fixed a bug which prevented additional quantizations from being downloaded for staff pick models that were already downloaded - Fixed a bug where `lms import` will sometimes not work properly if llmster (daemon) is also installed - Fixed a bug in `/api/v1/chat` that caused server errors when inputs were empty or `top_k` exceeded 500 - Fixed a bug where `lms ls` and `lms load` sometimes would fail after waking up the LM Studio service - Fixed a bug where sometimes token counting would not work properly for gpt-oss models **Build 15** - Introduce Parallel Requests with Continuous Batching 🚀 - When loading a model, you can now select n_parallel to allow multiple requests to be processed in parallel. - When enabled, instead of queuing requests one by one, the model will process up to N requests simultaneously. - By default, parallel slots are set to 4 (with unified KV set to true, which should result in no additional memory overhead). - This is supported for LM Studio's llama.cpp engine, with MLX coming later. - Introducing Split View in Chat: view two chats side by side. - Drag and drop chat tabs to either half of the window to split the view. - Close one side of the split view with the 'x' button in the top right of each pane. - Introducing 🔧 Developer Mode: a simplification of the previous Developer/Power User/User 3 mode switch. - Developer Mode combines the previous Developer and Power User modes into a single mode with all advanced features enabled. - You can turn on Developer Mode in Settings > Developer. - New setting: enforce allowing only one new empty chat at a time (default: enabled) - Change in Settings > Chat - New 🔭 Model Search experience - Access via the 🔍 button on the top right or by pressing Cmd/Ctrl + Shift + M - Model format filter preferences persist between app restarts - Modal is resizable and remembers its size between app restarts - Limit number of open tabs to 1 per pane. Support showing 2 side-by-side chat tabs. - Selecting a new chat replaces the current tab in that pane. - Add button to create a new chat in the sidebar - Pressing Cmd/Ctrl + L while the model picker is open will dismiss it - On narrow window size show right hand sidebar as an ephemeral overlay - Support for the LFM2 tool call format - CLI now uses commit hash for versioning instead of semantic version numbers - Updates to UI details in hardware settings - Fixed a bug where moving large number of conversations would sometimes only move part of them - Fixed a bug where `lms ls` sometimes would show incomplete list of models on startup - Fixed a bug in deleting tool confirmation preferences in settings - Fixed a UI bug in app onboarding - Fixed a visual bug in Models Table selected row affecting the Architecture and Format columns - Fixed a bug where undoing pasted content in chat input would not work as expected - Fixed a bug where a leading decimal in a numeric input would parse as a 0 - Fixed a bug rendering multiple images in a conversation message - Fixed a bug where a documentation sidebar section would sometimes get stuck in expanded state - Fixed a bug where chat names would sometimes be empty - Fixed a visual bug in rendering keyboard shortcuts on Windows and Linux - Fixed a bug where model loader would sometimes close due to mouse move shortly after opening - Fixed a bug rendering titles in preset conflict resolver dialog - Fixed a bug where reloading with new load parameters would not apply next time the same model is used for a chat - Fixed a bug where the model loading will get stuck if the cpu moe slider is maxed out - Fixed a bug where exporting chats with very large images to PDF would fail - Fixed a responsive UI overlap bug in the app header - [Windows] Fixed a bug where the default embedding model will not be available after in-app update - Adds download, copy, and reveal in working directory buttons to generated images in chat **Build 14** - (Build 14 was skipped) **Build 13** - App setting to control primary navigation position: 'top' or 'left' - [Mac] New tray menu icon 👾 (experimental, might change) - `/api/v1` endpoints and `/v1/responses` API now return better formatted errors - Significantly reduce the size of the app update asset **Build 12** - Bugfix: New chats to be created with the same model as the previously focused chat - Bring back gear button to change load parameters for currently loaded model - Bring back context fullness indicator and current input token counter - New in My Models: right-click on tab header to choose which columns to show/hide - New in My Models: Capabilities and Format columns - Fixed a flicker in model picker floating panel upon first open - P.S. you can open the model picker from anywhere in the app with Cmd/Ctrl + L - Fixed focus + Enter on Eject button not working inside model picker - Updated chat terminal and messages colors and style - Fixed dragging and dropping chats/folders in the sidebar **Build 11** - ✨👾 Completely revamped UI - this is a work in progress, give us feedback! - [CLI] New `lms chat` experience! - Support slash commands, thinking highlighting and pasting larger content - Slash commands available: /model, /download, /system-prompt, /help and /exit - [CLI] New: `lms runtime survey` to print info about available GPUs! - FunctionGemma support - Added a slider to control n_cpu_moe - New REST API endpoint: `api/v1/models/unload` to unload models - Breaking change: in `api/v1/models/load` endpoint response, introduced in this beta, `model_instance_id` has been renamed to `instance_id`. - Display live processing status for each loaded LLM on the Developer page - Prompt processing progress percentage → token generation count - Improved PDF rendering quality for tool requests and responses - Significantly increased the reliability and speed of deleting multiple chats at once - Updated style of chat message generation info - Updated layout of Hardware settings page and other settings rows - Fixed a bug where sometimes models are indexed before all files are downloaded - Fixed a bug where exporting larger PDFs would sometimes fail - Fixed a bug where pressing the chat clear hotkey multiple times would open multiple confirmation dialogs - Fixed a bug where pressing the chat clear hotkey would sometimes duplicate the chat - Fixed a bug where pressing the duplicate hotkey on the release notes would create a glitched chat tab - Fixed a bug where `lms help` would not work - Fixed a bug where deleting models or canceling downloads would leave behind empty folders - Fixed a styling bug in the GPU section on the Hardware page - [MLX] Fixed a bug where the bf16 model format was not recognized as a valid quantization **Build 10** - (Build 10 was skipped) **Build 9** - (Build 9 was skipped) **Build 8** - Fixed a bug where the default system prompt was still sent to the model even after the system prompt field was cleared. - Fixed a bug where exported chats did not include the correct system prompt. - Fixed a bug where the token count was incorrect when a default system prompt existed but the system prompt field was cleared. - Fix a bug where sometimes the tool call results are not being added to the context correctly - Fix chat clearing with hotkey (Cmd/Ctrl + Shift + Option/Alt + D) would clear wrong chat - Fix a bug where Ctrl/Cmd + N would sometimes create two new chats - Updated style for Integrations panel and select - Fixed cURL copy button for embedding models displaying additional incorrect requests - Fix "ghost chats" caused by moving conversations/deleting conversations **Build 7** - Fix jinja prompt formatting bug for some models where EOS tokens were not being included properly - Bring back release notes viewer for Runtime available update - Prevent tooltip from staying open when hovering tooltip content - Fix a bug in deleting multiple chats at once - Minor fix to overlapping labels in model loader - Support for EssentialAI's rnj-1 model **Build 6** - Fixed a bug where Qwen3-Next user messages would not appear in formatted prompts properly **Build 5** - Fixed a bug where quickly deleting multiple conversations will sometimes soft-lock the app - Fixed another bug that prevented the last remaining open tab from being closed **Build 4** - Fixed a bug where the last remaining open tab sometimes could not be closed - Fixed a bug where `lms log stream` would exit immediately - Fixed a bug where the server port would get printed as [object Object] - Image validation checks in `v1/chat` and `v1/responses` REST API now run without model loading - Fixed a bug where images without extensions were not classified correctly - Fix bug in move-to-trash onboarding dialog radio selection where some parts of the label were not clickable - Fix several clickable areas bugs in Settings windows buttons - Fixed a bug where certain settings may get adjusted unexpectedly when using llmster (for example, the JIT model loading may become disabled) - New and improved Runtime page style and structure - Fixes a bug where guardrail settings were not showing up in User UI mode - Fixed a bug where `lms log stream` would exit immediately **Build 3** - Introducing 'llmster': the LM Studio Daemon! - True headless, no GUI version of the process that powers LM Studio - Run it on servers, cloud instances, or any machine without a graphical interface - Load models on CPU/GPU and serve them, use via `lms` CLI or our APIs - To install: - Linux/Mac: `curl -fsSL https://lmstudio.ai/install.sh | bash` - Windows: `irm https://lmstudio.ai/install.ps1 | iex` - Support for MistralAI Ministral models (3B, 8B, 13B) - Improved `lms` output and help messages style. Run `lms --help` to explore! - Get llama.cpp level logs with `lms log stream -s runtime` in the terminal - `lms get` interactive mode now shows the latest model catalog options - New and improved style for Downloads panel - New and improved style for App Settings - We're trying something out: Model search now in its own tab - still iterating on the UI for this page, please give us feedback! **Build 2** - Show release notes in a dedicated tab after app updates - Add support to display images in exported PDFs and exported markdown files - Quick Docs is now Developer Docs, with refreshed documentation and direct access from the welcome page. - Allow creating permission tokens without allowed MCP permissions - Fixed a bug where sometimes images created by MCPs are not showing up - Fixed a bug where sometimes the plugin chips not working - Fixed a bug where the "thinking" blocks will sometimes expand erroneously - Fixed a bug where certain tabs would not open correctly - Fixed a bug where sometimes the model list would not load - Fixed a bug where in-app docs article titles would sometimes wiggle on scroll - Fixed a visual bug in Preset 'resolve conflicts' modal - Fixed a bug where sometimes Download button would continue showing for an already downloaded model - Fixed a bug where chat sidebar buttons wouldn't be visible on narrow screens - Display model indexing errors as buttons rather than hints **Build 1** - Welcome to the 0.4.0 Beta!