Qwen3.8-Flash-Next-MLX-4bit-Slotpack

Slotstream lossless download package

This repository contains the losslessly compressed download representation of pipenetwork/Qwen3.8-Flash-Next-MLX-4bit, for the Slotstream downloader. It reconstructs the exact original model files, without changing precision.

Use Slotstream to download and verify this package. These transport objects are not directly loadable safetensors. The original model remains available in the separate uncompressed mirror.

The manifest and every compressed object are content-addressed. Slotstream pins the repository commit and independently verifies compressed objects, reconstructed ranges, and every final original file. Interrupted pulls resume.

Credit for the conversion belongs to pipenetwork. The underlying model is Qwen/Qwen3.8-Flash-Next and remains subject to its Qwen community license. Slotstream's software license does not replace the model license.

See the download format guide for the format, exact byte counts, publication, and verification tools.

Decode forecast correction (sidecar)

lookahead/tap-correction-attention-rank128-v1.safetensors (37,540,708 bytes, SHA-256 37b00d3a32d1e1889a1794bbb8e97905a157a77c0508db620c1a11f2a895f7f5) is an optional file outside the Slotpack objects. It holds the rank-128 FP16 correction of the decode lookahead's router forecast that Slotstream 0.2.19 and later apply by default: slotstream pull fetches it after the weights, pinned by size, digest and repository commit, and a model directory without it runs the previous forecast. It is derived from this checkpoint's own routing on Slotstream's benchmark requests and changes no model weight and no output token. Its measurement and reproduction are in the expert lookahead guide.