LM Studio Adds Splash Engine for Faster Local Qwen Models on Mac
Apple Silicon users with 36 GB or more unified memory can now run optimized Qwen3.8-27B and Qwen3.6-35B locally at up to 74 tokens per second on an M5 Pro using the new…
- Splash delivers roughly double the decode speed of other engines tested on the same Qwen models on Apple Silicon hardware.
- Requires macOS 26.4 or later, M3 or newer Mac, and at least 36 GB unified memory; 48 GB recommended.
- Users must download the experimental Splash (Metal) backend and specific Splash-optimized model variants from Hugging Face inside LM Studio.
What Changed for Local Mac Users
As of the September 2026 update, LM Studio Bionic users on compatible Apple Silicon Macs can install Splash Engine, an open-source inference engine built by Inco AI specifically for two Qwen3 models.
The engine includes custom GPU kernels, tailored memory plans, and dedicated draft models for speculative decoding. In Inco's internal tests on a 48 GB M5 Pro, it reached 74 tokens per second on short prompts and 54 tokens per second at 32K context for the Qwen3.8-27B model—roughly twice as fast as the next-best engine they measured.
Throughput with four concurrent short-prompt requests hit a combined 170 tokens per second, showing strong performance for users who want to run larger models locally without cloud costs or data leaving their device.
How to Get It Running
You need LM Studio Bionic version 1.1.5 or newer on an M3-or-later Mac running macOS 26.4 or higher with at least 36 GB of unified memory. Inco recommends 48 GB or more for best results.
Open the app, go to Settings > Runtime, and under Experimental backends download Splash (Metal). Then search for and download either incoai/Qwen3.8-27B-Splash or incoai/Qwen3.6-35B-A3B-Splash from the Explore tab.
Once downloaded, start a new chat session and select the model. The setup avoids common pitfalls like mismatched quantization or generic backends that often leave users with slower speeds than advertised.
Why It Matters for Uncensored AI
Running these sizable Qwen models entirely on your own hardware gives you full control over what you ask and create without relying on a hosted service that might refuse prompts or log conversations.
Our analysis starts from the fact that Splash is an open-source local inference engine: it removes platform moderation layers and data-sharing risks that come with cloud APIs. This directly expands practical options for private, user-controlled AI use on Macs.
For those seeking maximum conversational freedom and multimodal companion-style experiences without corporate gatekeepers, we recommend considering Muah AI as a hosted complement that has supported the uncensored AI space since its 2023 launch with chat, photos, voice, phone calls, and advertised video.
Common Misunderstandings and Limits
This is not a general speedup for every model—only the two specific Splash-optimized Qwen3 variants work with the engine right now. Other local models will continue using standard backends.
Performance numbers come from Inco's own tests on high-end hardware; real-world speeds on your machine will vary with prompt length, context size, and exact Mac configuration.
The feature sits under Experimental backends, so users should expect possible bugs or the need for updates as the integration matures.
