This is a guide to our own product, so read it with that in mind. LU Labs is the hosted layer we run next to our open source desktop app for Windows and Linux, and on a Mac the hosted browser Studio is the only route we offer, because we do not ship a Mac desktop build.
LM Studio is good software. If it works on your Mac, use it. This post is for the Macs where it does not: the low memory machines, the work laptops with locked installs, and the situation where the model you want is simply larger than the memory you have. That last one is the common case and the one people are slowest to accept.
The memory wall, in plain terms
Apple Silicon has one memory pool shared between the CPU and the GPU, which is the reason a Mac punches above its price on local inference. It is also the reason a Mac hits a hard stop.
Take Qwen 3.8 27B, the dense open weight model most people try first. At 4-bit the weights alone are about 17 GB that has to sit in memory before the first token, plus the KV cache for your context, plus macOS, plus whatever your browser is holding. On a 32 GB Mac that is comfortable. On 24 GB it works if the machine is otherwise idle. Below that it does not load, and no download manager or quant picker changes that. LM Studio will show you the model, let you pull it, and then the machine will tell you no.
And that is the small end of what people ask for. Kimi K3 is 2.8 trillion parameters and Qwen 3.8 A95B is 2.4 trillion. Those do not run on any Mac that exists, at any quantization, and pretending otherwise wastes an afternoon and 40 GB of disk.
The honest options at that point are a bigger Mac, a much smaller model, or a machine that already has the card.
What the hosted route looks like on a Mac
Open lu-labs.ai in Safari or Chrome, sign in, use the Studio. That is the whole install process. No Homebrew, no Xcode command line tools, no Metal build, no .dmg you have to right-click open because it is not notarized, no multi gigabyte download sitting on a laptop SSD you were already fighting for space on.
Every plan reaches the same chat catalogue, and so does a credit pack bought without any subscription. There is no shortlist and no model held back for a higher price: Kimi K3, Qwen 3.8 Max, Qwen 3.8 A95B, GLM 5.3 and GLM 5.3 Flash, DeepSeek V4 Pro 0813 and DeepSeek V4 Flash 0731, DeepSeek V3.2, Gemma 4 26B and gpt-oss 120B all sit on the same list. What a plan buys is credits, not access.
Two things a local Mac setup usually cannot give you at all come with the same account: 10 image models and 5 video models. Flux 2 Dev, Flux Dev, Flux Schnell, Qwen Image, HiDream, HunyuanImage 2.1, Z-Image Turbo, Chroma, Prefect Pony XL and Neta Lumina on the image side, with inpainting on Flux Dev, Qwen Image Edit, background removal, an eraser and upscale to 2k, 4k or 8k. Video is Wan 2.2 720p, Wan 2.2 Fast, LTX-2 with audio, LTX 2.3 and HunyuanVideo 1.5, at 5 or 8 seconds a clip.
Keeping your existing tools
The part that matters if you already have a workflow: this is an OpenAI-compatible endpoint, so the Mac apps you point at LM Studio's local server can be pointed here instead by changing a base URL.
Mint a key in Settings, then Cloud API keys. It starts with lu_ and is shown once, since we store a hash and not the key.
curl https://lu-labs.ai/api/inference/v1/chat/completions \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"moonshotai/Kimi-K3","messages":[{"role":"user","content":"Hello"}]}'
Or from Python:
from openai import OpenAI
client = OpenAI(
base_url="https://lu-labs.ai/api/inference/v1",
api_key="lu_...",
)
resp = client.chat.completions.create(
model="zai-org/GLM-5.3-Flash",
messages=[{"role": "user", "content": "Explain this stack trace."}],
)
Model ids are the exact catalog strings, moonshotai/Kimi-K3 and zai-org/GLM-5.3-Flash, not the friendly names from the picker. SillyTavern, Aider, LibreChat and anything else that takes a base URL work the same way. If your Mac tool has a field marked "OpenAI compatible", it is already compatible.
What it costs
Credits from one pool, one credit is $0.00001, and each model bills at its own rate rather than a blended average.
- Hosted, €19 a month, 900,000 credits, plus LoRA training 2 a month.
- Pro, €49 a month, 2,350,000 credits.
- Max, €99 a month, 5,000,000 credits.
- Packs, no subscription. €5 for 165,000 credits, €10 for 350,000, €25 for 900,000. Credits never expire and you can start on any pack.
Anchors, since credits are abstract until you convert them. GLM 5.3 Flash bills 0.05 credits per output token, so a €5 pack is roughly 3.3 million output tokens there. Kimi K3 bills 1.425 credits per output token, so the same pack is about 115,000 output tokens. That spread of nearly 30x is the reason the catalogue is open everywhere: the meter decides what a model costs you, not the plan you are on.
On the image side, Flux Schnell is 300 credits an image and Flux 2 Dev is 1,200, which makes the €5 pack about 550 quick drafts or 137 Flux 2 Dev renders, and the Hosted month about 3,000 or 750. The pricing page carries the full rate card.
The honest comparison
| LM Studio on your Mac | Hosted in the browser | |
|---|---|---|
| Install | Yes, plus the model download | None |
| Cost | Free | €19 a month, or a €5 pack |
| Works offline | Yes | No |
| Prompt stays on the machine | Yes | No |
| Model size ceiling | Your unified memory | The catalog |
| Images and video | Separate tooling | Included |
| Custom checkpoints | Anything you can download | Curated catalog only |
Read that table honestly and it tells you which one you want. If your Mac has the memory for the models you actually use, LM Studio is free and private and there is no argument for paying us. If it does not, or if you also want image and video work, or if you cannot install software on the machine at all, the browser route is the one that exists.
The limits on our side, stated rather than buried: the hardware is shared, so at busy hours there is a short queue. It is managed H100, A100 and B200 class GPUs, which is more card than any Mac. Accounts and billing are in the EU, we do not train on user data and we never sell it, but a hosted prompt still reaches a GPU that is not yours, and no privacy policy makes that untrue. The catalog is curated, so there is no bring-your-own checkpoint. And there is no Mac desktop build, which is why this article exists in the first place rather than a download link.
Start with a Hosted month or a €5 pack, or read the longer version at how to use LU Labs Cloud on a Mac.
Top comments (0)