Qwen3.8-27B Splash 8-bit (unofficial)

A Splash package of Qwen3.8-27B with an 8-bit target model (GGUF Q8_0), for the Splash runtime in LM Studio on Apple silicon. It pairs the 8-bit weights with the same DFlash2 draft, vision tower, tokenizer and chat template as the official 4-bit package incoai/Qwen3.8-27B-Splash.

Unofficial and independent. Not made or endorsed by Inco AI (Splash, DFlash2), LM Studio, Alibaba Cloud (Qwen) or Unsloth. No weights were changed: every weight file is a byte-for-byte copy of its source.

Why 8-bit

The official package's target is 4-bit. 8-bit weights stay much closer to the original model, at roughly 60% of the 4-bit package's writing speed (see Measurements).

What's inside

Requirements

• Apple silicon M3 or newer, macOS 26.4 or later (Splash's requirements).
• LM Studio with the Splash runtime enabled (Settings → Runtime → Splash). Tested with LM Studio 0.4.26+4 and the Splash runtime 1.3.0.
• About 32 GB of memory for the model when loaded; tested on a 64 GB Mac. Download size: 34.8 GB.

Install

In LM Studio, enable the Splash runtime, then download dmeight/Qwen3.8-27B-Splash-8bit from the model search. It appears as Qwen3.8 27B Splash 8bit.

Known issues

• LM Studio labels it "4bit". LM Studio currently gives every Splash package a fixed 4-bit label, whatever its weights. The weights here are 8-bit: the package is 34.8 GB and takes about 31.5 GB of memory when loaded, while the official 4-bit package is 17.4 GB.
• It relies on undocumented behaviour, so an update may break it. LM Studio lists a folder as a Splash model only if its manifest.json names the 4-bit package format (splash-packed-q4). The Splash engine reads an installer assembly (model.json) before a manifest. This package carries both: the manifest for LM Studio, the assembly for the engine. A newer LM Studio or Splash runtime may stop loading it.
• It is a stopgap. Once LM Studio ships Splash 1.3.1 or later, which loads 8-bit models directly, prefer the official route.

Without LM Studio

This package is for LM Studio only. Elsewhere, use the original sources:

• Standalone Splash rejects this package, because its installer expects the official package's per-layer files. It doesn't need the package either: splash serve --model unsloth/Qwen3.8-27B-GGUF:Q8_0 builds the same set (the same GGUF, vision projector and DFlash2 draft) from the original repositories. Splash 1.3.1 and later also load MLX 8-bit (mlx-community/Qwen3.8-27B-8bit).
• llama.cpp, Ollama and other GGUF tools can use target/Qwen3.8-27B-Q8_0.gguf and vision/mmproj.gguf as they are: plain GGUF files, identical to Unsloth's. They run without the DFlash2 speedup. It is simpler to download them from unsloth/Qwen3.8-27B-GGUF.
• MLX tools cannot use it: the weights are GGUF, not MLX.

Measurements

Mac Studio, Apple M5 Max (40-core GPU), 64 GB, macOS 27.0.1, 2026-10-26.

Speed through LM Studio. Mixed prompts (Polish and English prose, code), thinking on, temperature 1.0, top_p 0.95, top_k 20, 800-token answers, two runs each, median.

Draft acceptance. The share of DFlash2 proposals the target accepted. 24 prompts in six groups, 512-token answers, ABBA order (4-bit, 8-bit, 8-bit, 4-bit), counters read from the Splash engine.

With sampling, the 8-bit target accepts more of the draft's proposals: +3.7 points on average over the 24 prompts, 95% bootstrap interval +1.1 to +6.5. Greedy decoding shows no difference. Greedy outputs were identical across repeated runs.

Licenses and credits

Every part is licensed under Apache-2.0 (see LICENSE); NOTICE lists the sources and what this package adds.

• Qwen3.8-27B by Alibaba Cloud (Qwen team).
• GGUF quantization and vision projector by Unsloth.
• Splash engine, package format, DFlash2 draft model and the official package's tokenizer and vision files by Inco AI.

Packaged by dmeight.