license: other
license_name: qwen-community-license-1.0
license_link: LICENSE
base_model: Qwen/Qwen3.8-Flash-Next
base_model_relation: finetune
pipeline_tag: image-text-to-text
library_name: finchmoe
tags:

• finch
• finchmoe
• qwen3.8
• flash-next
• moe
• quantized
• abliterated
• vision

Qwen3.8-Flash-Next 125B — Finch 2-bit + vision payload (abliterated)

This repository holds a .finch install for the out-of-core FinchMoE engine and apps — not a
Transformers checkpoint. It is the abliterated Qwen3.8-Flash-Next checkpoint repacked with 2-bit routed
experts, with the source vision tower carried along.

• Text source: windowsxp811203/Qwen3.8-Flash-Next-Abliterated
  at revision deb02632504bb214702bc28b0381a93d3112f500 (BF16 snapshot).
  Base model: Qwen/Qwen3.8-Flash-Next.
• Text body: resident weights int4/int8 affine (group 64); routed experts 2-bit affine
  (group 64, BF16 scales/biases); the frozen PLE n-gram table is int4-affine.
• Vision: the source model.visual.* tower — 333 BF16 tensors, 897,862,112 bytes
  (0.836 GiB) — copied byte-for-byte into vision/vision.safetensors, with the source
  config.json and both preprocessor configs beside it.
• Runtime status: the current FinchMoE text engine ignores the vision/ files. They are
  declared in the manifest, downloaded and verified like any other file, and reserved for the
  upcoming vision runtime and for other tools. Text generation works today.

Artifact facts

File layout

manifest.json                     file set, sizes, SHA-256, architecture, quantization
model_weights.bin                 resident weights (int4/int8 + fp16/bf16 runtime entries)
packed_experts/layout.json        expert tensor map
packed_experts/layer_XX.bin       one routed-expert pool per layer
ple_shards/shard_XXX.bin          quantized PLE n-gram table parts
tokenizer/                        tokenizer.json, config, merges, vocab, chat template
vision/vision.safetensors         source vision tower, BF16, unchanged
vision/model_config.json          verbatim source config.json (vision_config + token ids)
vision/preprocessor_config.json   image preprocessor configuration
vision/video_preprocessor_config.json  video preprocessor configuration

There is deliberately no verified-install.json here. That receipt records the physical
directory it was generated for, so a published copy would not validate elsewhere; the installer
writes its own after verifying every file against manifest.json.

Memory requirement

This is a 125B-class model. On the engine's default 16 expert-cache slots the routed-expert
cache alone costs 48 layers × 16 × 1.47 MiB ≈ 1.10 GiB, on top of the 3.88 GB resident
payload, KV cache and scratch. It fits 12–16 GB devices; an 8 GB iPhone cannot run it. For those,
use the Qwen 3.6 Finch builds instead.

Download and verification

The Finch app/CLI fetches a .finch distribution with its manifest: every file is requested at
its declared size and verified against its declared SHA-256, and an interrupted download resumes
against the same repo/commit. The manifest is the authority for the file set; vision/ files are
ordinary manifest entries.

Provenance and license

Repacked from the BF16 abliterated checkpoint above. The abliteration is upstream's work; see its
model card for the method, eval numbers and caveats. The vision tower, lm_head and n-gram table
are byte-identical to upstream. License: Qwen Community License 1.0 (see LICENSE), inherited
from the base model.