Qwen3.8-27B on Apple silicon with Splash: setup, requirements, and the verdict boundary

Splash is not a different name for LM Studio, and it is not a model. It is an experimental inference engine from Inco AI that LM Studio exposes inside the separate Bionic app. The supported model list is intentionally narrow: Qwen3.8-27B and Qwen3.6-35B-A3B.

Official requirements

These are platform floors, not a promise that every 36GB Mac will feel interactive. Unified memory is shared with macOS and other apps, and the model plus draft model plus KV cache all consume the same pool.

Set up Splash in Bionic

1. Install LM Studio Bionic 1.1.5 or newer. Do not confuse it with the classic LM Studio desktop app.
2. Open Bionic and go to Settings → Runtime.
3. Under Experimental backends, click Download next to Splash (Metal).
4. Go to Settings → Explore, paste one of the official Hugging Face repository links, select the model, and download it: https://huggingface.co/incoai/Qwen3.8-27B-Splash or https://huggingface.co/incoai/Qwen3.6-35B-A3B-Splash.
5. Start a new session and select the downloaded model from the local model picker.

Memory planning without inventing numbers

The official floor is 36GB, with 48GB recommended. That is the only memory guidance we can state without measuring the Splash artifacts ourselves. The draft model and KV cache make the real footprint larger than the model weights alone; how much larger depends on context and the engine's plan.

• If you have a 36GB Mac, close other memory-heavy apps before loading and expect less context headroom.
• LM Studio attributes its published Splash measurements to Inco's tests on a 48GB M5 Pro. That result does not establish performance on other Macs, even ones with 48GB or more.
• A larger Mac does not guarantee a speed tier: our engine has no Splash calibration.
• If the model fails to load, check the macOS version, Bionic version, engine download, and that you selected the exact Splash repository.

What we can compute for Apple silicon

For models that are in our dataset — including Qwen3.6-35B-A3B — the regular checker computes fit at your context using the quant artifact sizes in our dataset. Those are GGUF-style artifacts, not Splash's engine plan, so do not transfer the number to Splash as if it were measured.

• Check a tracked model on an Apple silicon machine: /can-i-run.
• Compare Mac Studio configurations on memory and price: /compare.
• Read the Ollama guide if you want the MLX preview path in Ollama instead: /runtime/ollama.

What this guide verified

We read LM Studio's Splash guide and Bionic announcement on 2026-10-19. We did not run Splash on an M3 Mac, and we did not measure tokens per second. The performance figures in LM Studio's post are attributed to Inco's tests, not to TokenAssemble.