Qwen3.8-27B-Splash

Qwen3.8-27B-Splash

Serves Qwen3.8-27B on Apple silicon Macs through Inco AI's Splash engine, with a 4-bit model, a DFlash 2 draft and a vision encoder in one package. Works only inside Splash.

Made by incoai 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 is Qwen3.8-27B packed for Splash, an open-source inference engine from Inco AI built for Apple silicon. The package holds the 4-bit model, a DFlash 2 draft model for faster decoding, a vision encoder and the tokenizer. Transformers and MLX cannot load it, so Splash is the only way to run it.

Who it suits: Mac owners and developers who want a 27B model running locally, behind an API that their coding agents and apps can call.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local model for coding agents on a Mac | A local assistant behind a terminal agent
IT operations | Small teams | Serve a private model from one Mac | An internal chat server
Research | ML researchers | Measure local inference speed on Apple silicon | A decode and prefill timing test
Content work | Writers and analysts | Ask questions about images and PDFs without a cloud service | A summary of a PDF kept on the machine

If you do not have a recent Apple silicon Mac with plenty of memory, you can skip it, since it needs at least 36 GB.

Where to use it:

• Macs with an M3 chip or newer, running macOS 26.4 or later
• A terminal, through the splash command
• Coding agents such as OpenCode, Claude Code, Codex and Hermes, which splash can attach

Use case | Who | What they get back
Local chat server | Developers | An API that speaks OpenAI Chat Completions, Responses and Anthropic Messages
Agent backend | Agent users | A local model behind a coding agent
Image and PDF input | Analysts | Answers about images and inline PDFs

How to set it up:

• A Mac with an M3 chip or newer and macOS 26.4 or later
• Homebrew
• 36 GB of unified memory, with 48 GB or more recommended
• About 17.4 GB of disk for the first download

1. Install Splash with Homebrew.
2. Start the server and name this model.
3. Wait for the Ready line, which appears after the download, check and memory test.

brew install incoai/tap/splash
splash serve --model incoai/Qwen3.8-27B-Splash

How to use it: After the server reports Ready, open the local page it prints in a browser, or attach an agent you already have from another terminal. Reasoning is on by default at the xhigh level.

splash opencode    # or: splash claude / splash codex / splash hermes

Tool | What it does
splash serve | Starts the local server for this model
splash opencode | Attaches OpenCode, with claude, codex and hermes as other options
reasoning_effort | Takes low, medium, xhigh or none
--max-memory | A ceiling setting for memory use
--max-context | A ceiling setting for context length

Example: You start the server, wait for Ready, then ask in the browser page, "Explain speculative decoding in one sentence". The model answers in text, with its reasoning returned separately as reasoning_content.

For newcomers and experts: Start with the two commands and the browser page, and try one short question. The server is open on your Mac only, with no authentication by default, so set an API key before you expose it further.

The package has a 14.1 GiB 4-bit target, with one packed file per layer. It also holds a 1.2 GiB DFlash 2 draft and a 0.9 GiB bf16 vision encoder. A manifest records a SHA-256 for every file. On an M5 Pro the card reports 74 tok/s decode on a short prompt, and a cached replay of a 32K prompt reaching first token in 282 ms.

Install Splash with Homebrew
Run splash serve with the model
Package downloads and is verified
Server reports Ready
Apps and agents call the API

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: downloads the package on first run, and the server answers on your own Mac only
• Not for: use outside Splash, since the package does not load in Transformers or MLX
• Not for: Macs below the stated chip, macOS and memory needs

Title: Qwen3.8-27B speed measured on an M5 Pro
Decode, short prompt: 74 tok/s
Prefill, 32K prompt: 363 tok/s
Decode, 4 concurrent short prompts: 170 tok/s

On Hugging Face, 13 November 2026

• Task: text generation
• Library: splash
• Downloads in the last 30 days: 0
• Likes: 187
• License: apache-2.0
• Based on: mlx-community/Qwen3.8-27B-4bit
• Tags: splash, apple-silicon, metal, local-inference, dflash2, speculative-decoding, qwen3.8, 4-bit, text-generation

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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