LM Studio

Updated November 2026: LM Studio now centers on Bionic, an agent for documents, code and automation, and it has added paid US-hosted cloud models alongside the free local setup.

How it works in practice

LM Studio is no longer just a model launcher with a chat window. The homepage now leads with Bionic, an agent that creates and edits documents, slides and PDFs, handles coding and automation, and can take over computer control tasks. Every change to a document saves automatically, so you can let the agent work without babysitting a version history.

The workflow splits into three paths. You download an open model inside the app and run it on your own machine through the LM Studio runtime, which sits on MLX and llama.cpp. Or you point Bionic at a bigger machine you own through LM Link. Or you pay for US-hosted open models when your hardware can't hold them. The older pieces still sit underneath: the OpenAI-compatible local server, the lms CLI, the JavaScript and Python SDKs, and the llmster headless runtime for Linux boxes and CI. If you used LM Studio as plumbing for scripts, none of that went away.

Voice is the other addition. Bionic transcribes speech in real time, in multiple languages, and the vendor says audio is processed on your device and never leaves it. The free plan lists offline transcription. For anyone who dictates notes and can't send recordings to a third party, that's a concrete reason to try it. If transcription is your main need, our roundup of open-source Otter AI alternatives that run on your own machine covers dedicated tools.

The homepage download button reads "x64". Check that a build exists for your operating system and chip before you plan around it, especially on a Mac.

What stands out

The free tier keeps the agent. Most local-model apps give you a chat box and charge for anything agentic. Here the Bionic agent, local models through llama.cpp and MLX, and offline voice transcription all cost $0. That's a stronger free offer than the cloud agents it competes with, as long as your hardware can run a model good enough to drive an agent. A weak local model makes a weak agent, and no app fixes that.

Open frontier models, with a stated no-retention policy. The homepage names GLM 5.2, Kimi K3 and DeepSeek V4 Pro as the models to run for demanding work, and the paid plans list Kimi K3, GLM 5.3 and DeepSeek V4 Flash. The vendor says its cloud services use Zero Data Retention, meaning prompts and responses aren't stored, and its FAQ says the same for web search. It also warns that web content is untrusted and that you should watch what goes into queries. That caveat matters more for an agent that reads pages and then acts than for a chatbot. Treat ZDR as a vendor statement to weigh, not an audit result. The vendor lists GDPR as its compliance mark and says customer content isn't used for training.

LM Link makes the hybrid sensible. It connects your devices so a laptop can use models loaded on a desktop or server elsewhere. The free plan covers up to 5 devices. If you own one strong GPU machine and a thin laptop, you get large local models on the laptop without paying for hosted inference. That's a smarter first move than upgrading to a paid plan.

The developer surface matters if you build things. Agents increasingly talk to tools over Model Context Protocol, and LM Studio supports it as a client. Read our explainer on what MCP is and what can go wrong before you give a local agent access to files and apps.

Where it falls short

Hardware sets the ceiling. The runtime is efficient, but a model that doesn't fit in your memory won't run at a useful speed, and CPU-only inference stays impractical beyond small models. Model downloads are large, so budget disk space before you collect several.

The free plan caps web search at "limited", and the vendor doesn't publish the number. If your agent work depends on research across the web, you'll hit the paid tier or look elsewhere. Page extraction comes with Bionic+, not the free plan.

Bionic is new. An outside review from Kompozy, last verified in July 2026, scored it 4.0 out of 5. It praised local execution, privacy, coding and document work, but called the product brand new and noted that it generates no media and publishes nothing. That matches the vendor's own framing. This is a work agent over your code and files, not a content factory. The changelog shows 7 entries in the last 90 days, the latest on 21 October 2026. That's a healthy pace, and it also means behavior will shift under you. Pin versions if you script against the CLI or SDKs.

Teams are the weakest spot. Centralized billing exists through inference credits, and you can create an organization and invite people, but subscription plans for teams are listed as coming soon. Enterprise pricing isn't published, so a department that wants predictable per-seat billing today can't get it from the pricing page.

Value for money

The free plan is the product for most people. Local models, the agent, offline transcription and 5-device LM Link cost nothing, and you only pay in hardware and electricity. If you already own a capable GPU or a recent Mac, there's little reason to pay.

Bionic+ at $20 a month makes sense in one situation: you want frontier-scale open models such as Kimi K3 or GLM 5.3 and can't host them. It also adds web search, page extraction and discounted bulk tokens. At that price it's cheaper than most paid cloud assistants, but you get open models, not the closed ones many people compare against. Judge the output on your own tasks first, using the free tier with a local model.

Pro at $100 a month buys 5× the usage limits and early access to new features. The public plan summary doesn't state the base allowance, so the multiplier is hard to turn into real work. Take Pro only if you've repeatedly hit the Bionic+ limits. Both prices are flat per account rather than per seat, and cloud credits cover usage beyond what a plan includes. Heavy users should expect to watch that meter.

Who should pick it, who should skip it

Pick it if you handle sensitive documents or code and want an agent that runs on your own hardware, if you already own a strong GPU or Apple Silicon Mac, or if you want one app that runs local models for free and lets you buy hosted open-model capacity only when you need it. Developers who already use the OpenAI-compatible server or SDKs get the agent on top at no extra cost.

Skip it if your machine is a low-end laptop, if you need guaranteed per-seat team billing today, or if your work depends on the strongest closed cloud models. Pure command-line users who don't want a desktop app should look at Ollama, which stays the simpler fit for scripts and servers.

Setup notes

Start on the free plan with a mid-size local model, and test your real task before paying for anything. If you move to cloud models, remember that web search results are untrusted input to an agent, so keep sensitive text out of queries.