MikeZ75/Qwen3.8-Flash-Next-V3-Splash

Qwen3.8-Flash-Next V3 — Slipstream package (Splash Q4)

nitinpanj/qwen38-flash-next-v3, converted into the ready-to-run package format of Slipstream, the C++/Metal inference engine for Apple Silicon. Slipstream serves the GGUF too, but converts it on its first start; this package skips that step and needs no disk space for a second copy.

This is not a GGUF. It is Slipstream's splash-packed-q4-qwen4exp format (schema 5) and runs only in Slipstream. For llama.cpp, use the original GGUF linked above.

For UkisAI's Swift variant of this model, see MikeZ75/Swift-Qwen3.8-Flash-Next-V3-Splash.

Use

curl -fsSL https://github.com/mzinner/slipstream/raw/main/install.sh | sh
slipstream serve --model MikeZ75/Qwen3.8-Flash-Next-V3-Splash

The first start downloads the package into ~/.slipstream/models and verifies every file against manifest.json. Or use Slipstream Menubar, which installs Slipstream and the model and runs the server from the menu bar.

Requirements: an Apple Silicon Mac with 64 GB of memory, and about 100 GiB of free disk.

Contents

Part	Files	Size
Target model: 48 layers, embedding, head, n-gram table, MTP layer	target/*.bin	~99.8 GiB
Draft model (a placeholder the format requires)	draft/*.bin	~0.1 MiB
Tokenizer	tokenizer/	~20 MiB
Every file with its size and SHA-256	manifest.json

Routed experts and the n-gram table are tiled 4-bit (groups of 64); attention, mixers, shared experts, embedding and head are 8-bit. The MTP draft head is the one shipped with the GGUF, and target/draft-vocab.bin comes from nitinpanj/Swift-Qwen3.8-Flash-Next-Splash (both models share the vocabulary).

Converted with Slipstream's GGUF converter (models/qwen4exp/tools/convert_qwen4exp_gguf.py).

Credits and license

• Model: Qwen3.8-Flash-Next by Qwen (Qwen/Qwen3.8-Flash-Next).
• GGUF, MTP head and draft vocabulary: nitinpanj, who also wrote the Slipstream runtime and its converter (npanj/slipstream).
• This conversion: MikeZ75. No weights were changed beyond the format conversion.

Licensed like its source, under the Qwen Community License 1.0.