As of July 2026, the latest version is LM Studio 0.4.18, released June 26, 2026. This page tracks the LM Studio 2026 updates from 0.4.10 through 0.4.18: MCP OAuth support, Qwen 3.6, the Locally mobile app, LM Link leaving its waitlist, multi-GPU tensor parallelism and MTP speculative decoding. It gets updated as new versions ship.
Over those months LM Studio got faster on multi-GPU machines, reached phones, and learned to sign in to authenticated MCP servers.
Release notes for 2026
Newest first.
| Version | Date | Changes |
|---|---|---|
| 0.4.18 | Jun 26, 2026 | Bug-fix release: fixed reasoning on/off being ignored by some models, zoom/clipping glitches, and chat-title generation for reasoning models. |
| 0.4.17 | Jun 25, 2026 | LM Studio Engine Protocol now off by default; Mermaid diagrams open fullscreen and export to PNG; AMD Strix Halo support via llama.cpp 2.22.1. |
| 0.4.16 | Jun 4–8, 2026 | Launched Locally, LM Studio's mobile app; LM Link no longer needs a waitlist; default context bumped to 8k tokens; multi-GPU selection fixes. |
| 0.4.15 | May 29, 2026 | CUDA tensor parallelism for multi-GPU loading; physical batch-size option; fixes for Claude Code and codex --oss. |
| 0.4.14 | May 22, 2026 | Stable MTP speculative decoding, which speeds up generation with models that ship built-in multi-token prediction heads. |
| 0.4.13 | May 13, 2026 | MLX engine v1.8.1 with parallel predictions for vision models on Apple Silicon; chat input paste fixes. |
| 0.4.12 | Apr 17, 2026 | Added Qwen 3.6 support; fixed a Windows bug affecting OAuth-backed MCP servers. |
| 0.4.10 | Apr 9, 2026 | Added OAuth support for MCP servers. |
Tensor parallelism (0.4.15) and MTP speculative decoding (0.4.14) make multi-GPU and capable single-GPU setups faster. The mobile app and LM Link (0.4.16) stretch LM Studio past a single desktop, so one strong machine can serve lighter ones. MCP OAuth, from April, is what made authenticated tool use workable in the first place, and the later releases build on it.
What changed in April
Locally AI joins LM Studio (April 8)
LM Studio said it had acquired Locally AI, an app for running your own AI models on iPhone, iPad and Mac. Its creator, Adrien Grondin, joins the team to lead work on native AI experiences across devices.
The fit is a setup plenty of people already have: a bigger machine (Mac mini, Mac Studio, a gaming PC) does the heavy inference and lighter devices act as clients. LM Studio's LM Link documentation describes the same idea. You access your local models across devices over an end-to-end encrypted connection, running large open-weight models on a powerful machine and using them from a lighter laptop as if they were local. Locally AI adds a focus on iPhone, iPad and Mac.
Not every detail has been announced, so the direction is an inference, not a promise. The likely shape is a personal model network: the heavy box runs the model, and a phone, second laptop or travel machine is the interface.
MCP OAuth in 0.4.10 (April 9)
LM Studio 0.4.10 added OAuth support for MCP servers. A server that requires sign-in no longer forces you into copied tokens, manual headers and brittle config notes.
Many services don't work with static tokens. They use sign-in flows, approvals, refreshed credentials and scoped access, and OAuth gives LM Studio a cleaner path to them. You are also less likely to leave credentials in text files or shell history, and there are fewer custom hacks to maintain.
A local model gets more useful once it can reach documentation tools, coding systems, internal dashboards or productivity services, the same kind of authenticated services people already use in cloud agent workflows. If you edit mcp.json by hand, a JSON formatter or validator catches typos, and a diff tool shows what differs between a working config and a broken one after an update.
Qwen 3.6 in 0.4.12 (April 17)
The 0.4.12 release notes added support for Qwen 3.6 and fixed a Windows bug where OAuth-backed MCP servers would not work in some environments. The same release improved chat PDF export styling and Qwen 3.5 performance across compatible API endpoints.
Setting up a cross-device workflow
- Update LM Studio to 0.4.12 or newer, which gives you Qwen 3.6 support and the Windows OAuth fix.
- Pick one model for one real task. A compact instruct model suits writing and summarising, Qwen 3.6 suits coding or reasoning, and a model you already know behaves well with tools suits MCP. Don't begin with five.
- Run the heavy inference on the strongest machine you own: a desktop GPU box, a Mac Studio, or a server-class machine at home.
- Reach that model from your lighter device with LM Link. LM Studio says it lets you use large models from a stronger machine as if they were local, so a laptop or secondary Mac no longer carries the full compute load.
- Add one OAuth-backed MCP integration that helps a real task: a documentation or search workflow, a coding tool, or a file or knowledge base.
- Test through the API or your usual coding tool. If your editor, API client or agentic coding tool talks to the model as though it were local, the setup works. For the API side, see our follow-up guide: LM Studio Local Server Setup: Run a Private OpenAI-Compatible API.
- Keep a short note of your default model, your preferred inference machine, the MCP integrations you rely on, and what changed after each update. It saves hours later.
Example setups
- Laptop plus desktop: a stronger desktop at home holds the heavier model. A travel laptop connects through LM Link for writing, code review and questions, without hosting the big model itself.
- Mac and iPhone: this is where the Locally AI deal pays off, if LM Studio keeps pushing deeper Apple-device integration. The appeal would be a private, personal workflow that follows you across devices you already own.
- Authenticated tools: connect one OAuth-backed MCP integration and let the model query a service that needs a login. That is when local AI stops feeling like an offline toy.
FAQ
What is the latest LM Studio version in 2026?
0.4.18, released June 26, 2026, as of July 2026. It is mainly a bug-fix release. The larger recent feature updates were 0.4.16 (the Locally mobile app and LM Link leaving the waitlist), 0.4.15 (CUDA tensor parallelism) and 0.4.14 (stable MTP speculative decoding).
Does LM Studio support MCP OAuth?
Yes. LM Studio 0.4.10 added OAuth support for MCP servers, and 0.4.12 fixed a Windows issue affecting some OAuth-backed MCP server setups.
Is this only news for Apple users?
No. The Locally AI acquisition strengthens the Apple side, but OAuth-backed MCP support, LM Link, the Windows fixes and support for current open models apply more widely.
Should I switch to LM Studio?
Consider it if you want local models, cross-device access and authenticated tool use. If you only want the easiest possible AI experience, cloud tools may still be simpler.
Sources
- LM Studio blog: Locally AI joins LM Studio, published April 8, 2026
- LM Studio changelog: version 0.4.10, published April 9, 2026
- LM Studio changelog: version 0.4.12, published April 17, 2026
- LM Studio changelog: full release history including 0.4.13 through 0.4.18 (latest June 26, 2026)
- LM Studio docs: Using MCP via API
- LM Studio docs: LM Link
