mxcake3893/Qwen3.8-27B-Alis-Splash

Qwen3.8-27B-Alis (Splash packed Q4)

Splash-format model package for Qwen3.8-27B-Alis (pre-graded), converted from the MLX 4-bit safetensors release into the native Splash packed layout.

Two base repositories are referenced:

• incoai/Qwen3.8-27B-Splash — the official Splash package this one is built against (identical layout; draft/, tokenizer/, and vision/ are taken from it)
• avlp12/Qwen3.8-27B-Alis-MLX-4bit — source of the target weights (pre-graded, snapshot f9a9e88)

⚠️ Compatibility: use the original Alis with current Splash

This packed model is for older Splash only. Current Splash no longer supports this format. Alis itself is still a supported model architecture—use its original MLX repository instead.

Your Splash version	Which model to use
Older build: f43509a	This repository — confirmed working with our local llama-server/Splash worker integration
Current Splash	Original Alis MLX — the upstream migration path

Run Alis on current Splash

splash serve --model avlp12/Qwen3.8-27B-Alis-MLX-4bit

Want the source weights used by this packed release? Pin the original revision. The source repository's latest weights and quantization have changed:

splash serve --model avlp12/Qwen3.8-27B-Alis-MLX-4bit \
  --revision f9a9e88e2326d4089a866463fd6a499d14c75862

Speed and memory after migration: this repository's uploader has not measured the difference between the old packed release and current Splash loading the original MLX model. Unchanged speed or memory usage is not guaranteed.

Why did this stop working? / Exact compatibility commits

Splash removed its legacy splash-packed-* loader on October 6, 2026, in commit 539eae9. It now loads upstream MLX/GGUF checkpoints directly.

The confirmed local deployment is based on Splash commit f43509a2f03d83b4442a0fab748e8b0fb49aa817, with the local llama-server/Splash worker integration. This does not certify every later commit or a separately tested, unmodified upstream CLI.

Package

• Format: splash-packed-q4, schema 3 (4-bit, group 64, StorageN=256 affine transform)
• Lossless conversion: the MLX 4-bit quantized values are preserved exactly; the pipeline was verified to reproduce the official incoai/Qwen3.8-27B-Splash byte-for-byte (66/66 section SHA256 match) from the matching base model mlx-community/Qwen3.8-27B-4bit before this fine-tune was converted.

Accuracy — quoted from the original Alis README

The accuracy results below are quoted from the original model's README, not measured by this repository's uploader. They describe the source pre-graded MLX build used for this conversion—not an independent evaluation of this packed package or of current Splash.

Source: avlp12/Qwen3.8-27B-Alis-MLX-4bit README at the source revision f9a9e88. The figures and evaluation conditions are reproduced below; table labels are shortened for readability.

Results reported in the source README: full-vocabulary KL against bf16 (248,320-way softmax, ≈100K scored tokens).

Source build	KL to bf16 (nats/token)	Top-1 agreement
4-bit AWQ pre-graded	0.0654	91.3%
Uniform 4-bit (mlx-community)	0.0763	90.5%
NVFP4 (mlx-community)	0.0962	89.0%
MXFP4 (mlx-community)	0.1437	86.4%

Source-reported strided perplexity: context 2048, ≈103K scored tokens.

Dataset	Source pre-graded build	BF16 reference	Excess
English (wikitext-2)	5.8450	5.7734	+1.2%
Korean (Korean Wikipedia)	6.2609	6.0954	+2.7%
Code (CPython)	1.8105	1.6813	+7.7%

"It is also the build with a real quality cost — the excess over bf16 is significant on every slice: smallest on English, about twice that on Korean, largest in relative terms on code. Read the quality section before you deploy it."

— Original Alis README, linked above.

For interpretation and deployment recommendations, consult the original README. This repository does not present these quoted results as its own measurements.

Contents

Path	Description
manifest.json	Package manifest (schema 3), with per-artifact size + SHA256
target/	66 binary sections: packed Q4 weights, bf16 scales/biases, embeddings, head logits, GDN decay
draft/	Draft model (copied from the official package)
tokenizer/	Tokenizer files
vision/	Vision tower weights

Usage

Serve with the Splash native Metal runtime:

splash serve-native ./target ./draft <max_context|auto> <max_memory|auto>

or through the llama.cpp server integration (splash worker), which accepts this package as a model root.

Credits

• Fine-tune / quantization / measurements: avlp12 (Qwen3.8-27B-Alis-MLX-4bit)
• Base model: Qwen3.8-27B (mlx-community/Qwen3.8-27B-4bit)
• Package format reference: incoai (Qwen3.8-27B-Splash)