@lmstudio founder • lmstudio.ai/jobs
New York
- Tweets 1,231
- Following 832
- Followers 4,883
- Likes 6,242
I had the huge privilege of presenting at WWDC this year, demoing @lmstudio's upcoming clustering feature live on stage at the Steve Jobs Theater.
This was only possible because of the incredible work of the legendary MLX team, present and past: @trebolloc, @angeloskath , @awnihannun and everyone else.
Thank you for having me there! 🍎👾
10
10
133
16,759
OMFG Karpathy saying LM Studio is his favorite for running local LLMs in his new vid! 🤯🙇♂️
(but also spends a full minute ripping on the old UX 😭😭😭)
Sensei, you need to update! We fixed all of that in 0.3. Though always lots more room for making things simpler and more accessible.
Thank you so much for the shoutout @karpathy! <3
New 3h31m video on YouTube:
"Deep Dive into LLMs like ChatGPT"
This is a general audience deep dive into the Large Language Model (LLM) AI technology that powers ChatGPT and related products. It is covers the full training stack of how the models are developed, along with mental models of how to think about their "psychology", and how to get the best use them in practical applications.
We cover all the major stages:
1. pretraining: data, tokenization, Transformer neural network I/O and internals, inference, GPT-2 training example, Llama 3.1 base inference examples
2. supervised finetuning: conversations data, "LLM Psychology": hallucinations, tool use, knowledge/working memory, knowledge of self, models need tokens to think, spelling, jagged intelligence
3. reinforcement learning: practice makes perfect, DeepSeek-R1, AlphaGo, RLHF.
I designed this video for the "general audience" track of my videos, which I believe are accessible to most people, even without technical background. It should give you an intuitive understanding of the full training pipeline of LLMs like ChatGPT, with many examples along the way, and maybe some ways of thinking around current capabilities, where we are, and what's coming.
(Also, I have one "Intro to LLMs" video already from ~year ago, but that is just a re-recording of a random talk, so I wanted to loop around and do a lot more comprehensive version of this topic. They can still be combined, as the talk goes a lot deeper into other topics, e.g. LLM OS and LLM Security)
Hope it's fun & useful!
piped.video/watch?v=7xTGNNLP…
53
88
2,074
244,189
Desktop apps are HARD!
Glad to finally release LM Studio (lmstudio.ai)
If you're curious about running open-source LLMs on your laptop ... but prefer a GUI over running CLI commands -- LM Studio might be for you.
Built on llama.cpp from @ggerganov & the community
19
95
578
172,405
Finally got a mac with more than 16GB RAM. Although I remain a 7B believer, being able to run a 32B model locally and having it be always on feels like a totally different ballgame
36
4
458
55,897
LM Studio 0.3.8 will ship later today with Thinking UI, LaTeX rendering improvements, and bug fixes
16
22
422
28,699
LM Studio is now free to use anywhere: both at home and at work.
I put the original terms in place back in August 2023, shortly after launching the app. The idea was to allow anyone to experiment with LM Studio at home for free, but require companies to get in touch and buy a commercial license.
This made sense at the time, but soon added way too much friction to teams to adopt LM Studio at work.
Today we're removing this requirement and making it free to use LM Studio at home, work, school, or anywhere.
The blog post also shares info about how we'll make money! Let us know your thoughts and feedback.
LM Studio is now free for use at work.
Starting today, it no longer necessary to get a separate license to use LM Studio for work. No need to fill out a form or contact us. You and your team can just use the app!
LM Studio has always been free for personal use. This is due to our fundamental belief that AI should be accessible to people on their own machines, with no reliance on an external party, and with complete privacy.
The purpose of this change is to realign the terms with our mission of making local AI accessible and useful, and make it easier for teams to adopt LM Studio at work (or at school, or anywhere).
Nothing else changes: the app still does not snoop on you, no account is required, and your data always remains private and local to your machine.
10
25
382
29,916
We've integrated Apple MLX into LM Studio!
MLX is Apple's new on-device machine learning framework. It's extremely fast and flexible.
Read more about it: lmstudio.ai/blog/lmstudio-v0…
7
39
320
21,927
gpt-oss-20b + LM Studio's built-in code interpreter (runs sandboxed JS using Deno) is a useful combo
9
19
258
24,605
We just released LM Studio 0.3.10 with Speculative Decoding support!
It's a technique that pairs together a large model, and a more efficient (smaller) "draft" model that can generate tokens fast (draft generation). The larger model then verifies/rejects the draft tokens. If enough tokens are accepted, you sometimes get very significant speedups, could be on the order of 2x+ in some cases.
Some good combos:
- Llama (8B/70B) and Llama 1B
- Qwen (14B/32B/72B) and Qwen (0.5B/1B)
- Any other model family that has a spectrum of sizes
Would be cool to see more small models coming out with the explicit goal of being draft models for larger ones.
In LM Studio's chat / server you can easily switch up draft models in the sidebar and see what works best. There are many optimizations still to make to get things even faster!
Our implementation integrates llama.cpp work from @ggerganov and the llama.cpp community, as well as MLX work from @awnihannun and the MLX community. LM Studio's own @mattjcly really cooked with this one 🔥.
Let us know how you like it!
LM Studio 0.3.10 is here with 🔮 Speculative Decoding!
This provides inferencing speedups, in some cases 2x or more, with no degradation in quality.
- Works for both GGUF/llama.cpp and MLX models!
- Easily experiment with different draft models
- Visualize accepted draft token % rate
- Works in Chat UI and API
13
31
259
18,847
llama.cpp developers and community came together in a really impressive way to implement Qwen3-VL models. Check out the PRs, it’s so cool to see the collaboration that went into getting this done.
Standard formats like GGUF, combined with mainline llama.cpp support ensures the models you download will work anywhere you choose to run them.
This protects you from getting unwittingly locked into niche providers’ custom implementations that won’t run outside their platforms.
🎉 Qwen3-VL is now available on llama.cpp!
Run this powerful vision-language model directly on your personal devices—fully supported on CPU, CUDA, Metal, Vulkan, and other backends.
We’ve also released GGUF weights for all variants—from 2B up to 235B. Download and enjoy! 🚀
🤗 Hugging Face: huggingface.co/collections/Q…
🤖 ModelScope: modelscope.cn/collections/Qw…
📌 PR: github.com/ggerganov/llama.c…
5
17
229
31,180
openai/gpt-oss models are now in LM Studio!
These are the best local models we've ever run. The 20B is wicked smart. This is going to change a lot of things.
and they're Apache 2.0
Meet gpt-oss, OpenAI's open models!
openai/gpt-oss-20b
openai/gpt-oss-120b
Download and run them now on Mac, Windows, or Linux with LM Studio!
10
14
187
22,153
Such an awesome moment! LM Studio shown in Apple's M4 MacBook pro announcement. It's in the video too 🥹
LM Studio featured in Apple's M4 MacBook Pro announcement!
> With support for up to 128GB of unified memory, developers can easily interact with LLMs that have nearly 200 billion parameters.
16
10
176
16,833
In the next version of LM Studio:
> Run in headless mode: use the underlying LLM server without the GUI (both GGUF and MLX)
> Just-In-Time model loading: sending an inference call will load the model on demand
Try the beta: lmstudio.ai/beta-releases
Let us know what you think!
14
30
175
15,577
That's on us, no auto update 0.2 -> 0.3 atm. Need to download from the website: lmstudio.ai/download
8
1
172
10,696
Excited to finally launch the LM Studio SDK for Python and TS/JS! The underlying goal is to enable people to use LM Studio's capabilities programmatically to build tools, scripts, apps, and more.
In addition to the standard APIs you'd expect (chat completions, embeddings), we also added an "agentic" API: the .act() call. It automates the process of calling tools -> providing the tool output to the model -> letting the model decide what to do next.
We get into all the details in the blog post, let us know what you think!
Announcing the LM Studio SDK for Python and TypeScript! 🔨
As a part of the SDK, we also introduce our first agent-oriented API: the `.act()` call.
You provide a task and tools, and the model goes on its own autonomously for multiple turns until it accomplishes the task.
8
21
169
14,767
The new @MistralAI coding model is really good a tool calling and you can run it locally on your Mac or PC.
Needs about ~15GB of GPU memory to run fast.
Devstral Small is now available in LM Studio!
This model achieves a score of 52.4% on SWE-Bench Verified, outperforming the previous Devstral Small by +5.6% and the second best state of the art model by +10.2%.
Please update to LM Studio 0.3.18 first. Enjoy!
4
9
170
10,610
New @lmstudio version with gpt-oss tool call parsing fixes + MLX support is on the way
7
4
173
9,620
Starting the next version of LM Studio (0.3.6) it's going to be a piece of cake to update your MLX or llama.cpp engine. Won't require a full app update. Excited for this!
8
11
150
10,538
Phew! LM Studio 0.3.0 with RAG, Structured Outputs API with any local LLM, and tons more features in finally here.
Incredibly proud of the team for the work on this release.
If you want to experiment with LLMs locally on your computer, give it a try!
Works on Mac, Windows (x86/arm), and Linux (x86).
After months of work, and with the help of our awesome community, we're excited to finally share LM Studio 0.3.0! 🎉
🔥 What's new:
- Built-in Chat with Documents, 100% offline
- OpenAI-like 'Structured Outputs' API with any local model
- Total UI revamp (with dark/light/sepia themes)
- Load & serve multiple LLMs *on the local network*
- Available in 7 languages! 🌎🌍🌏
- Download any supported model from Hugging Face
- Update LLM runtimes (llama.cpp) separately from the app
... and tons more goodies!
Let us know how you like it! 👾🤝
14
18
137
13,912
small model (<4b), long context (250K+), and super good at tool calling - who's building this?
19
2
131
6,772
'venvstacks' enables us to ship our Python based MLX engine with LM Studio without requiring users to install dependencies themselves.
Excited to open source it today and share how it works!
github.com/lmstudio-ai/venvs…
4
12
121
7,106
llama.cpp is the foremost enabler of on-device AI. It's also an astonishingly well run open source project
Technical blog by @nvidia about using llama.cpp on RTX-powered systems:
developer.nvidia.com/blog/ac…
"More than 50 tools and apps are now accelerated with llama.cpp on the RTX platform"
Nice! If you are one of these tools give us a shoutout :)
2
12
108
10,997
LM Studio founder here. We don’t make money yet. But we raised a round that enables us to fund operations. We’re working with early adopters for “LM Studio @ Work” which will involve some paid complement but it’s still very nascent. Let me know if you have any questions. Fan of your work btw, cheers
7
2
94
4,037
really amazing to see the @lmstudio team working right now. Small but mighty group. Putting in huge effort and care into a couple of features we can't wait to share with you
4
5
91
3,865
Google cooked with Gemma 3. The 27B QAT version is legitimately great and is currently my default choice. Excellent at instruction following and tool use
Easily download and run Gemma 3 QAT models on your desktop with @lmstudio! No code required to get started: goo.gle/4lLkudO
ALT Image shows the UI and information for LM Studio download of Gemma 3 QAT
6
8
87
7,956
In the next update, LM Studio will be able to download models from nested folders in HF repos.
Workaround: download the files via the browser and manually move them to ~/.lmstudio/models/unsloth/DeepSeek-R1-GGUF
4
7
86
6,981
In the next version of LM Studio, you can get `reasoning_content` separately from regular content in your API requests to local DeepSeek R1 (distilled or full)
4
4
83
5,949
I'm not sure how HF is paying for all those TBs going in and out, but at least now we're chipping in a little bit. Thanks @huggingface for being the great library of AI models for us all 🙏
5
7
79
15,186
Now updated with MLX quants too
deepseek-r1-0528-qwen3-8b is here! 🔥
As a part of their new model release, @deepseek_ai shared a small (8B) version trained using CoT from the bigger model.
Available now on LM Studio. Requires at least 4GB RAM.
4
77
5,959
ByteDance/Seed-OSS (36B) now works in LM Studio beta build 5.
It has this interesting feature that it tries to budget its own reasoning
2
1
78
3,403
qwen3-14b is really strong at tool calling. Easily handled ~5000 tok convo that involved multi step file manipulations (ls, read, find what to change, make edit). Suggested creating backups before editing some files. Can do CoT, tool call, back to CoT and then more tools
4
3
76
4,279
In addition to the venerable chat completions compat API, @lmstudio now supports /v1/responses!
1. Swap out the openai base url to point to LM Studio
2. Load up gpt-oss
3. Profit
Introducing OpenAI Responses API compatibility!
/v1/responses on localhost.
Supports stateful responses, custom tool use, and setting reasoning level for local LLMs.
👇🧵
5
5
71
9,137
Better together 🤝
LM Studio's new mlx-engine architecture is an effort to unify the brilliant work of @awnihannun @angeloskath (mlx-lm), @Prince_Canuma (mlx-vlm) and community contributors.
This pattern is designed to be extended and we have a good first issue on the repo! 🍎
Introducing a new multi-modal architecture in LM Studio's MLX engine.
It's an extendible pattern for supporting new modalities in MLX models.
mlx-engine is open source (MIT). Check out the technical writeup for details and info on how to contribute!
lmstudio.ai/blog/unified-mlx…
3
9
78
9,304
Early preview of LM Studio 0.4.0 👀
Kicking off Data Driven NYC no 117 in front of a packed house with a *never before seen* demo of the latest & greatest from @yagilb, Founder & CEO of @lmstudio
Stay tuned for the same best-in-breed local AI, but now with !! Tool Use !!, coming soon to a laptop near you
11
3
73
4,096
FWIW in the current version of LM Studio we return the reasoning content inside the regular content, which could explain why it wasn't good via API. We made a change to separate out the reasoning content. Building now, should be ready in ~1 hr. Will be LM Studio v0.2.22
4
1
63
8,077
gemma 3 27B (qat GGUF quant) is really good at tool calling, you should give it a try if you can. Also showing impressive capability to follow what’s going on in the conversation. It was able to get references to earlier tool requests and execute them again with some changes I asked for. Was also able to do some good problem solving when the tool returned an error (file already exists), which surprised me. Takes about ~18GB RAM for 4096 context on my macbook
5
3
64
5,991
If gpt-oss was bugging out for you during in-chat tool calling, try out the latest beta. We squashed a few more bugs
4
2
59
6,725
For sensitive or private data (personal or business), nothing comes close to the peace of mind and feeling of ownership you get running local models on your own hardware
7
6
59
7,902
Currently using existing MCP servers or cooking your own? Want to use them with local LLMs?
Would love to hear from you: docs.google.com/forms/d/e/1F…
4
7
56
6,098
Our MLX engine is MIT, and the latest version bump is imminent: github.com/lmstudio-ai/mlx-e…. BTW you can also always check the engine release notes with Cmd + Shift + R within the app (or peek inside ~/.lmstudio/extensions/backends).
2
3
59
22,019
🥲 first @lmstudio TechCrunch mention!
DeepSeek's distilled new R1 AI model can run on a single GPU | TechCrunch techcrunch.com/2025/05/29/de…
5
8
53
3,761
the original model weight are MXFP4 as well. That's the same fidelity you get with the llama.cpp/GGUF format. So MXFP4 is full fat for this specific model
1
49
1,707
To run LFM2 in @lmstudio, do this:
1. Make sure you’re on 0.3.18
2. Press Cmd / Ctrl + Shift + R
3. Choose the “Beta” channel
4. Update llama.cpp
5. Run LFM2!
Try LFM2 with llama.cpp today!
We released today a collection of GGUF checkpoints for developers to run LFM2 everywhere with llama.cpp
Select the most relevant precision for your use case and start building today.
huggingface.co/LiquidAI/LFM2…
1
6
45
3,753
LM Studio CLI
Introducing lms -- LM Studio's companion cli 😎
✨ Load / unload LLMs, start/stop the local server
📖 Debug your workflows with lms log stream
🛠️ Run `npx lmstudio install-cli` to install lms
🏡 MIT licensed, code on github
lms requires LM Studio 0.2.22 or newer
Here are few examples of what you can do with lms 👇
3
7
41
5,051
lms has some pretty useful commands:
lms ls # list all local models
lms ps # list all loaded models
lms load # load a model
lms server start/stop # manage the local server
lms get # download a model
And my personal favorite:
lms log stream # show the formatted prompt that goes to the model
Did you know that LM Studio ships with a CLI tool called lms?
We added docs for its subcommands and flags.
📚 Check it out! lmstudio.ai/docs/cli
1
37
4,237
The future of local llm builders in NYC is bright.
Hackers up and early at the @LMStudioAI NYU hackathon. This is who you are competing with
3
2
36
4,377
Reminder: on-device AI hackathon is happening this weekend (Dec 7-8) @ @nyutandon.
Apply on the luma page: lu.ma/41hfiu79.
(Spots are limited)
If you're around Manhattan next weekend & want to hack on local LLMs:
We're hosting a small builder focused hackathon for ~20 teams at NYU. Windows on ARM64 is a thing now, and LM Studio is getting some new APIs too 👀
2
6
35
5,602
Small docs update: you can now add ".md" to any lmstudio.ai/docs route to get the page in raw markdown
Useful for feeding SDK docs to an LLM
1
5
33
3,923
it's our fault 🥲 But he's on the latest version now!
ok, updated!
(probably the update message should tell you that 0.3 exists but you have to download it manually at [url]?)
0.3 does look nicer/cleaner.
I still think the Discover tab can be even more improved for an average person with simple/sensible recommendations.
Appreciate your work though and happy to feature briefly in the video!
2
33
4,110
We've fixed a bug in LM Studio's MLX engine with side-loading models you've quantized yourself. This should work now. Version update coming tomorrow, but you can try it in beta already.
0.3.4 build 4: lmstudio.ai/beta-releases
Please let us know if this fixed these issues for you, and if you're running into any others!
1
4
30
4,379
One of my favorite new features in LM Studio 0.3.6:
"Cmd/Ctrl + Shift + J" to pop open engine logs (e.g. llama.cpp).
Useful for seeing what's happening under the hood and debugging models crashes / OOMs
1
1
25
2,982
We've added headless mode and on-demand model loading to LM Studio v0.3.5.
Much more ergonomic for development, and frees up lots of system resources if you're only using the server.
`lms server start`
LM Studio v0.3.5 is here! 👻🎃🥳
- Headless mode (run as local LLM service)
- On-demand model loading
- Support for @MistralAI's Pixtral with 🍎MLX
- `lms get` to download models from the terminal
lmstudio.ai/blog/lmstudio-v0…
2
1
25
2,244
MLX + MoE = 🚀
👏👏👏 @awnihannun @angeloskath
.@AIatMeta's Llama 4 Scout ripping 42.27 tok / sec on M2 Ultra Mac Studio
1
25
4,016
.@mattjcly fixed an mlx-engine bug that caused undesired prompt reprocessing. The difference in performance is huge, and the fix is pretty interesting. Check out the PR to see how this issue related to repetition penalty: github.com/lmstudio-ai/mlx-e…
Prompt caching can be deceptively tricky!
LM Studio MLX engine v0.14.0+ (right) fixes a bug in <=v0.13.2 (left) where follow-ups to long excerpts could unnecessarily force a full prompt re-compute 😅
Fix is out now in LM Studio v0.3.15.
Try it out @ lmstudio.ai/download
1
25
1,581
Mistral is so back?
.@MistralAI's Small 3 in now available in LM Studio! 🎉
- 24B params
- Apache 2.0
- Over 81% accuracy on MMLU
From the terminal, run: "lms get mistral-small-24b"
22
1,292
We've added support for @deepseek_ai R1 distilled models in LM Studio 0.3.7 (both llama.cpp and MLX!). These models represent a significant capability leap for open models. We'll be adding UI to hide/show content between <think></think> tags next.
DeepSeek R1 Distilled models are now available in LM Studio!
1.5B, 7B, 8B, 13B, 14B, 32B, and 70B variants.
Please update to LM Studio 0.3.7 first. Let's go! 🚀
2
1
22
1,026
Yes as far as I’m aware LM Studio is the fastest option right now, and will be even faster on Mac once we release the next MLX engine version (thanks to epic work from the MLX team)
19
461
LM Studio v0.1.1 is here! This release doubles down on making it easier, friendlier & more accessible for anyone to experiment with the latest local / offline LLMs.
All you need is a regular Mac or Windows PC. No command line or PhD required 😇
Demos videos & resources below.
2
7
18
6,186
Huge thanks to @awnihannun (MLX, Apple), @remilouf (outlines, @dottxtai), and @Prince_Canuma (mlx-vlm) for making this possible, and for helping us integrate MLX / Outlines into LM Studio 🙏
LM Studio's MLX engine is on github (MIT) github.com/lmstudio-ai/mlx-e…
1
17
1,089
If you’re in NYC next week come hang out with @lmstudio and @Letta_AI (of memGPT fame) teams and learn about stateful local agents
We're hosting the first 𝘚𝘵𝘢𝘵𝘦𝘧𝘶𝘭 𝘈𝘨𝘦𝘯𝘵𝘴 𝘔𝘦𝘦𝘵𝘶𝘱 in NYC on 𝗝𝘂𝗹𝘆 𝟭𝟳𝘁𝗵 🗽
Co-hosted with @lmstudio -- come by to learn about building stateful agents, fully locally
1
5
20
2,135
Even though qwen/qwen3-4b-thinking-2507 is a very small model it was able to execute a specific tool call flow I asked for. It thought for a while and then did it. Also interesting that it went into CoT again at the end before the final answer. No system prompt in this case, but there were a few simpler tool call requests earlier in the context.
20
1,231
Here's what running the WizardLM model (7B, q4_0, converted to ggml by @TheBlokeAI) on my 16GB RAM M1 MacBook pro.
You can paste this URL into LM Studio to run it too: huggingface.co/TheBloke/wiza…
4
1
17
4,553
@LMstudioAI v0.1.3 supports MPT (chat, instruct, storywriter) and StarCoder derived (e.g. WizardCoder) models.
I asked @WizardLM_AI's WizardCoder 15B to "Please generate a Typescript + React + Tailwind CSS web page for my flower shop business" and this is what it replied
1
5
17
3,353
Be the first product designer on the @LMStudioAI team.
Required: deep familiarity with computing lore
You’re a NYC product designer who can code and you believe the future of LLMs (and large models in general) is open source.
DM me your portfolio, I have a marvelous job opportunity. 📭
1
5
16
4,110
This is your reminder that the @lmstudio + @cline NYC AI community meetup is happening tomorrow (4/22).
We will be distributing never-before-seen LM Studio swag 👀
We're hosting our first developer meetup of the year!
Join the LM Studio team + @cline in NYC.
📆 Tuesday, April 22nd (Earth Day 🌎), 6:30PM
Agenda:
- Meet fellow builders
- Demos from @cline and @lmstudio
- Q&A and mingling
Register to attend ↓
1
2
15
2,939
"S2 is built on a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data."
Interesting.
Today, we're launching our first technical report!
The report details Helix, a generalist Vision-Language-Action (VLA) model
figure.ai/news/helix
3
16
1,344
this is fantastic news for open source model authors and users, love to see it 🦾
The Transformers library is undergoing it's largest pivot to date 🙌
It now cements its role as the central model definition, irrespective of the backend and runner.
One ground truth to bring more reliability across the ecosystem.
Why is this important?
15
879