Use Splash in LM Studio Bionic if you have an M3-or-newer Mac with macOS 26.4 or later and at least 36GB of unified memory. This guide follows Bionic's published setup for the 4-bit Qwen3.8-27B Splash package. Inco recommends 48GB or more. A 16GB or 24GB Mac should use a different setup for this particular package.
LM Studio added the integration on September 18. Splash is Inco AI's local inference engine, with model-specific kernels and a matching DFlash 2 draft model. The draft proposes tokens, and the main model checks them. That can reduce generation time, but it does not turn a large model into a small download.
Check the prerequisites first
| Requirement | What to check |
|---|---|
| Mac chip | Apple menu > About This Mac must show M3 or newer. An Intel, M1, or M2 Mac does not meet this Splash setup's requirement. |
| Operating system | macOS 26.4 or later. Check the version before downloading the model. |
| Memory | At least 36GB unified memory; 48GB or more recommended. Leave room for macOS and other applications. |
| Application | LM Studio Bionic 1.1.5 or newer, the separate agent app. |
| Disk and internet | The Qwen3.8 Splash package is about 17.4GB, plus the engine and working files. Leave extra space and complete the downloads first. |
If your Mac falls short, start with the Mac setup guide and its smaller-model path. The standalone Splash engine also documents smaller GGUF variants for some 24GB Macs; those are a different setup from Bionic's package recipe below.
1. Install Bionic and the Splash engine
Open the LM Studio download page and choose Bionic for your Mac. The original LM Studio application remains available on the same page; installing that app alone does not give you this Bionic settings path.
After opening Bionic:
- Go to Settings > Runtime.
- Find Experimental backends.
- Click Download next to Splash (Metal).
- Let the engine finish installing before selecting a model.
Download access uses the internet. For the first session, keep your project small and avoid adding remote services or web-search tasks.
2. Download the matching model package
Go to Settings > Explore, paste this full model link into the search bar, select it, and click Download:
https://huggingface.co/incoai/Qwen3.8-27B-Splash
This package includes the 4-bit target model, DFlash 2 draft, vision encoder, and tokenizer. It is a Splash package, not a general Transformers or MLX checkpoint. Selecting a similarly named model for another engine is a different installation.
The package's model card lists Apache 2.0 for its components. Use that package's terms when distributing weights; an engine's license by itself does not grant rights to every model it can load.
Bionic's instructions also offer incoai/Qwen3.6-35B-A3B-Splash as an alternative. Start with one package so you can identify which model and runtime your session actually uses.
3. Run a small first task
Once the download finishes, start a new session and choose Qwen3.8's downloaded package from the local model picker.
Paste this fictional example:
Turn these notes into three bullets. Do not add facts.
- The meeting starts at 10am.
- Jordan will bring the draft agenda.
- We need a decision about the launch date.
Check that the answer contains the three supplied facts, with no invented launch date or extra participant. This gives you a first result you can inspect without attaching real files or giving an agent permission to edit anything.
Then try a second message asking which decision remains open. The answer should identify the launch date. If the model adds a date, correct the prompt and compare the next response before expanding the task.
For document work, select the project's input files deliberately and review proposed changes. A locally running model can still call tools that access other files or the internet.
What the speed announcement means
Inco reports faster decoding and cached follow-up requests on its tested 48GB M5 Pro. The result depends on the model, hardware, prompt length, cache state, and number of concurrent requests. A cached repeat is a different workload from reading a new long document.
For your own comparison, use the same short task first, then the same longer text. Record the wait before output starts separately from the rate once generation begins. Do not buy a Mac on the assumption that an advertised maximum applies to every task.
If something goes wrong
| Symptom | Next step |
|---|---|
| Splash is missing from Runtime | Confirm you opened Bionic, meet the Mac/OS requirements, and installed version 1.1.5 or newer. |
| The model cannot load | Check the full repository name and that the engine download completed. This package needs Splash. |
| Memory error or heavy swapping | Close other loaded models and memory-heavy apps. Start a short new session; do not begin with a large document collection. Read the memory failure guide. |
| It downloads again or fails partway | Check free disk space and connectivity. Both model and runtime files must finish downloading. |
| The session uses a hosted model | Select the downloaded model in the local picker again before sending your next prompt. |
| A tool asks to search the web or edit files | Review that action separately. Choosing a local model does not grant a tool unrestricted access. |
Keep the application and model choices separate
Bionic supports local inference, LM Link, and hosted models. This recipe selects the local path. Web search, a linked remote computer, or cloud inference can introduce network traffic even with a local model installed.
Use the Bionic local-versus-cloud explanation to understand those choices. If you want a desktop chat workspace with broader runtime tuning, follow the original LM Studio installation guide. Once this short task works, the PDF guide explains extraction and retrieval before you ask questions about longer documents.