DeepSeek-V4-Flash-0731-abliterated-vision-v2

<i>After creating the our original fintune (https://huggingface.co/cbert33/DeepSeek-V4-Flash-0731-abliterated-vision), Drwozeys came out with a more narrowly abliterated version of the Deepseek model. We like this version better as it seems to be better with coding, but the personality is a little stiff. So you might like the previous version better. </br> We used the drowzeys version as the base and grafted on a vision tower that was under 1gb in size. Note that standard vLLM, as well as Eugr and Anemll don't fully support this setup, so we have a custom runner that does.</i>

DeepSeek-V4-Flash-0731 Abliterated Vision 2.0

This repository composes the language checkpoint from drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors with the DeepEncoderV2 tower and trained projector from FlyCockpit/DeepSeek-V4-Flash-0731-vision.

The operation is a component composition, not weight averaging or language-weight blending. The Drowzeys language-model shards remain byte-identical to their pinned source. The custom runtime loads the vision tower and projector from explicit paths and inserts projected image embeddings at the image token.

Uncensored model: the language checkpoint has undergone abliteration to reduce refusal behavior. Treat outputs as untrusted, apply application-level safeguards, and do not assume the model will decline harmful requests.

User responsibility: this model is provided without warranty. The creators, uploaders, and maintainers are not responsible or liable for what others generate, publish, deploy, or otherwise do with this abliterated model. Users must operate it responsibly, apply appropriate safeguards, comply with applicable law, and respect third-party rights. This model is for research purposes only and is not intended for production use.

Serving

This model is intended to pair with our dgx-spark-vllm-deepseek-v4-vision 2.0 for vision serving with vLLM on two NVIDIA DGX Spark systems. This provides the integration required for the DeepEncoderV2 tower, projector, image-token handling, and distributed vision path used by this model.

At the time of release, upstream vLLM, Eugr, and Anemll do not provide the complete vision-serving path required by this model (we loved those guys but had to make the update for our model). Use the companion runner and point it at the complete model directory to enable image-conditioned requests.

Example request shape:

{
  "model": "DeepSeek-V4-Flash-0731-Vision",
  "chat_template_kwargs": {"thinking": false},
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "Read the text in this image."},
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
    ]
  }],
  "max_tokens": 128
}

Language checkpoint

The 2.0 language component is Drowzeys' DSpark-compatible anchored-tensor checkpoint:

• Abliteration edits are limited to layers 10–35.
• The edited tensors are attn.wo_b weights with λ 3.5 and one refusal direction.
• Layers 36–42 are restored hash-identical to the official stock checkpoint.
• DSpark target layers 40–42 remain stock.
• MTP/draft tensors were not edited (edit_mtp: false). These properties are recorded in ABLIT_META.json and DSPARK_ANCHOR_MANIFEST.json and pinned in SOURCE_PINS.json.

Composition

1. Copy the complete pinned Drowzeys language checkpoint at revision a1e69379517383be9cd78c67defb04e77ad6aa68.
2. Preserve all 48 language shards and model.safetensors.index.json byte-for-byte.
3. Copy the DeepEncoderV2 tower and projector from FlyCockpit at revision d8efc7dfaceee965164d95952e2498b60fee323c.
4. Set the architecture to DeepseekV4VisionForCausalLM and add the explicit vision contract to config.json.
5. Add the pinned custom runtime artifact and model metadata.
6. Verify the complete output against MANIFEST.sha256.

Pinned identities

Machine-readable pins are in SOURCE_PINS.json. MANIFEST.sha256 covers every packaged file except itself.

Vision contract

• Architecture: DeepEncoderV2 tower, MLP projector, and learned view separator
• Input image size: 1024 × 1024
• Image token: <｜image｜> (129279)
• Tokens per view: 256
• Tiling: one global view plus up to four local crops from a 2×2 grid
• Image-token counts: 257, 769, or 1281, depending on aspect-aware tiling
• Qualified cache format: nvfp4_ds_mla

Validation and limits

The assembled Drowzeys + FlyCockpit artifact has been qualified through the custom runtime with text and image-conditioned requests. This is bounded deployment qualification, not a comprehensive visual, long-context, or safety benchmark. See VALIDATION.md.

• One image per prompt was qualified.
• OCR smoke tests do not establish broad visual-reasoning quality.
• The language checkpoint is abliterated and may generate unsafe, false, or policy-violating content.
• Maximum-context admission does not establish semantic quality at maximum length.

Reproducibility and notices

See ASSEMBLY.md for the composition and verification algorithm, SOURCE_PINS.json for immutable inputs, and THIRD_PARTY_NOTICES.md for attribution and licensing information.

Sources

1. Official DeepSeek-V4-Flash-0731
2. Drowzeys anchored-tensor language checkpoint
3. FlyCockpit vision assets
4. Vision plugin lineage