license: mit
library_name: transformers
pipeline_tag: image-text-to-text
base_model: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
tags:

• deepseek-v4
• multimodal
• vision
• abliteration
• preview
• draft

DeepSeek-V4-Flash-Vision-Exp Abliterated

[!WARNING]
PREVIEW / DRAFT RELEASE. This upload contains the initial native
abliterated checkpoint. Final behavioral testing, broader capability and
safety validation, runtime validation, and quantized GGUF/MLX variants will
follow. The recipe, metadata, and catalog layout may change before the final
release.

This is a model catalog. The directly loadable native checkpoint lives in
Reference-Native-FP8/; future GGUF or MLX profiles belong in descriptive
sibling folders and must be registered in variants.json.

.
├── Reference-Native-FP8/   native mixed FP8/FP4 model (48 shards)
├── README.md               catalog and usage notes
├── NOTICE                  source and direction provenance
└── variants.json           machine-readable variant registry

The native checkpoint uses a conservative transferred rank-1 direction and
changes only 33 attention output writers plus their FP8 scale
payloads. No weights from the abandoned
DeepSeek-V4-Flash-0731-Abliterated-Vision graft are used; only its useful
catalog organization inspired this layout.

Status: weight-transfer and byte-boundary validated; behavior not locally
validated. See the native folder's README and manifest for exact details.

Preview status

Completed for this draft:

• all 48 native shards and 72,633 indexed tensors are present;
• all 33 intended writer edits were deterministically replayed;
• exactly 66 FP8 writer/scale payloads changed;
• non-target payloads, vision modules, experts, routers, and MTP heads were
  verified against the official source;
• source, direction, recipe, and checksum provenance are included.

Still planned before the final release:

• end-to-end text and multimodal behavioral evaluation;
• refusal, benign-quality, regression, and safety testing;
• validation on compatible production inference runtimes;
• calibrated GGUF and MLX quantized variants with profile-specific testing.

This preview should be treated as an experimental research artifact, not as a
fully characterized or production-ready model.