license: other
license_name: qwen-community-license-1.0
license_link: https://huggingface.co/haihengh/Qwen3.8-Flash-Next-125B-finchmoe-4bit-ple4bit-abliterated-vision/blob/main/LICENSE
base_model: windowsxp811203/Qwen3.8-Flash-Next-Abliterated
pipeline_tag: image-text-to-text
library_name: finchmoe
tags:

• finch
• finchmoe
• qwen3.8
• qwen
• moe
• quantized
• vision
• image-text-to-text
• 4-bit
• abliterated

Qwen3.8-Flash-Next-125B Abliterated — FinchMoE 4-bit routed experts + vision payload

This repository holds a .finch install for the out-of-core
FinchMoE engine and its Mac/iOS apps —
not a Transformers checkpoint. It is the Qwen3.8-Flash-Next-125B checkpoint repacked with
routed experts at 4-bit, with the source vision tower carried along.

• Source: windowsxp811203/Qwen3.8-Flash-Next-Abliterated at the BF16 snapshot recorded below.
• Text body: attention / shared experts at 4-bit, linear attention and router at 8-bit, PLE table at int4 group 32; routed experts 4-bit affine
  (group 64, BF16 scales/biases).
• Vision: the source model.visual.* tower — 333 BF16 tensors,
  856.3 MiB — copied byte-for-byte into vision/vision.safetensors,
  with the source config and preprocessor sidecars. The FinchMoE vision runtime
  consumes it; text-only consumers can ignore it.
• Text identity: the text bytes are the same recipe as the published text
  sibling haihengh/Qwen3.8-Flash-Next-125B-finchmoe-4bit-ple4bit-abliterated; this
  repository adds the vision payload only.

Artifact facts

File layout

manifest.json                     file set, sizes, SHA-256, architecture, quantization
model_weights.bin                 resident weights
packed_experts/layout.json        expert tensor map (51.5 MiB)
packed_experts/layer_XX.bin       one routed-expert pool per layer
tokenizer/                        tokenizer.json, config, merges, vocab, chat template
vision/model_config.json                      4745 B  source vision tower and configs
vision/preprocessor_config.json                390 B  source vision tower and configs
vision/video_preprocessor_config.json          385 B  source vision tower and configs
vision/vision.safetensors                  856.3 MiB  source vision tower and configs

The repository carries the verified-install.json receipt generated while
verifying the published payload; the installer still re-verifies every file
against manifest.json.

Quality and performance

This variant is not separately measured. Its text weights are the same recipe
as haihengh/Qwen3.8-Flash-Next-125B-finchmoe-4bit-ple4bit-abliterated; use that card
for the measured quality and throughput of the text path. The vision tower adds
about 856.3 MiB of BF16 weights, mapped only while an image is being
encoded.

Usage

FinchMoECLI --model /path/to/Qwen3.8-Flash-Next-abliterated-ple4bit-vision.finch \
            --messages-file messages.json --max-new 128

The Mac/iOS app needs a vision-capable model row for image turns; a text-only
row will reject the vision/ directory rather than silently ignore it. Until
this repository is added to the app's picker, use the CLI or point the app at
the directory directly.

Provenance

Repacked with FinchMoERepack on 2026-10-28:

FinchMoERepack --input-snapshot <Qwen3.8-Flash-Next-125B-bf16> \
  --output Qwen3.8-Flash-Next-abliterated-ple4bit-vision.finch --routed-expert-bits 4 --include-vision

The source snapshot hash is recorded in the manifest above. The abliteration,
where applicable, was done upstream (windowsxp811203/Qwen3.8-Flash-Next-Abliterated); this repository changes
only the container, the quantization, and the added vision payload.

License

See LICENSE. Weights are derived from the upstream model named in
base_model; this repository is a derivative work.