Qwen3.8-27B Splash 8-bit

Qwen3.8-27B Splash 8-bit

Runs Qwen3.8-27B at 8-bit through the Splash runtime in LM Studio on Apple silicon. An unofficial package that keeps the model closer to the original than the 4-bit one, at lower speed.

Made by dmeight on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.8-27B Splash 8-bit is a community repackaging of Qwen3.8-27B for the Splash runtime that LM Studio offers on Apple silicon. The main model is a Q8_0 GGUF file, so the weights keep more precision than the official 4-bit package. A smaller draft model speeds up writing, and a vision tower lets it read images. It is independent of Inco AI, LM Studio, Alibaba Cloud and Unsloth, and none of the weights were altered.

Who it suits: Mac users who run local models in LM Studio and want 8-bit weights that stay closer to the original model, and who accept slower writing.

Industry | Who uses it | How they use it | Example
Software development | Developers on Apple silicon Macs | Run a 27B model locally in LM Studio with speculative decoding | Asking coding questions with the 8-bit weights
Writing and research | People who prefer fidelity to speed | Choose 8-bit over the 4-bit package | Comparing answers from the two packages on one prompt
Local AI enthusiasts | Mac Studio and laptop owners | Test image input through the vision tower | Giving the model a picture and a question

Skip it if you do not use LM Studio on an Apple silicon Mac, since the package is for that setup only.

Where to use it: Only in LM Studio with the Splash runtime enabled, on Apple silicon M3 or newer. The author lists macOS 26.4 or later.

• LM Studio, with the Splash runtime turned on in its settings
• A Mac with Apple silicon M3 or newer
• About 32 GB of memory when loaded, and a 34.8 GB download

Use case | Who | What they get back
Local chat | Mac users | Replies from the 8-bit 27B model
Image questions | Mac users | Answers that take a picture into account
Speed and accuracy comparison | Enthusiasts | A way to compare 8-bit with the official 4-bit package

How to set it up: Turn on the Splash runtime in LM Studio, then download the package from the model search.

• A supported Mac and a recent macOS
• LM Studio with the Splash runtime enabled
• Enough free disk space for the download

1. Open LM Studio and enable the Splash runtime in its settings.
2. Search for dmeight/Qwen3.8-27B-Splash-8bit and download it.
3. Load the model that appears as Qwen3.8 27B Splash 8bit.

You can also fetch the files with Python:

pip install -U huggingface_hub
hf download dmeight/Qwen3.8-27B-Splash-8bit

How to use it: Chat with it in LM Studio like any other model. Outside LM Studio the package does not work, and the card points to the original sources instead. Standalone Splash can build the same set from the original repositories with this command:

splash serve --model unsloth/Qwen3.8-27B-GGUF:Q8_0

Tool | What it does
LM Studio | Runs the package with the Splash runtime
splash serve | Builds the same set from the original sources in standalone Splash
llama.cpp and Ollama | Can use the plain GGUF files, without the draft model speedup

Example: You enable the Splash runtime, load the 8-bit package in LM Studio and ask a long writing question. It answers at roughly 55 tokens per second in the card's test, against about 91 for the official 4-bit package.

For newcomers and experts: If you are new to this, 8-bit means the weights keep more precision than 4-bit, at the cost of more memory and slower output. Try it in LM Studio, and check that your Mac meets the requirements first.

If you know the field, the target is a GGUF Q8_0 file and the draft is a BF16 DFlash2 model that Splash quantizes to 4-bit at load. The card reports the share of draft proposals the target accepts, 28.5 percent at 8-bit against 25.7 percent at 4-bit under sampling.

DFlash2 draft proposes tokens
8-bit Qwen3.8 target checks them
Splash accepts the matching ones
LM Studio shows the reply

Good to know: The card lists known issues. LM Studio labels the package as 4bit although the weights are 8-bit. The package depends on undocumented behaviour that an update may break. It is a stopgap until LM Studio loads 8-bit models directly.

Title: Card speed test, tokens per second when writing
Package | Tokens per second
This package, 8-bit | 54.8
Official Splash, 4-bit | 91.2
LM Studio MLX 8-bit | 18.1

• License: Apache 2.0
• Cost to run: none named, LM Studio and a suitable Mac are needed
• Data it reaches: the prompts you give it in LM Studio
• Not for: MLX tools, standalone Splash or other setups, since the package is for LM Studio only

On Hugging Face, 13 November 2026

• Task: text generation
• Library: splash
• Parameters: 27.3B
• Weights: about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 1,809
• Likes: 4
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Tags: splash, safetensors, gguf, qwen3_5, apple-silicon, metal, local-inference, dflash2, speculative-decoding, qwen3.8, 8-bit, lm-studio

Links

• Model: https://huggingface.co/dmeight/Qwen3.8-27B-Splash-8bit
• Files: https://huggingface.co/dmeight/Qwen3.8-27B-Splash-8bit/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download dmeight/Qwen3.8-27B-Splash-8bit

Then follow the usage in its model card. Nothing in this kit runs by itself.