drowzeys / keys-DeepSeekV4Flash-Vision-EXP-ablit

keys-DeepSeekV4Flash-Vision-EXP-ablit

Abliterated DeepSeek-V4-Flash-Vision-Exp . Public repo, gated with automatic approval after you agree to RESPONSIBLE_USE.md .

0731-style wo_b projection on L10–35 (λ=3.5, 26 tensors). L0–9, L36–42, MTP/DSpark draft, and the full vision tower stay stock.

See METHOD.md . Access is gated with automatic approval after you agree.

What was edited

Keys 0731 lesson: smash early layers or the in-checkpoint drafter and DSpark accept dies. Vision-Exp adds a vision tower on top of 0731 text. This dest:

• does project residual wo_b on L10–35
• does not edit MTP / DSpark draft tensors
• does not touch vision.* , aligner.* , image markers, or ffn.gate.bias_vl

Artifacts: ABLIT_META.json .

Vision vs this vLLM serve

The weights include the Vision-Exp tower (~309 vision tensors). Current Anemll DSpark vLLM ( DeepseekV4ForCausalLM ) is text-only and skips those tensors at load, including on the DSpark drafter (shared checkpoint, no vision params). Sending an image to that serve returns is not a multimodal model .

Native pixels on this checkpoint need a multimodal vLLM class for DeepSeek-V4-Flash-Vision, not a different model (Aeon / Qwen3.8 Omni is a separate stack). The designed workaround on the DSpark recipe is an optional Qwen3-VL sidecar ( ENABLE_VL_SIDECAR=1 , :8889 ), which does not use this tower.

Download

# after you agree to the gate (automatic approval)
hf download drowzeys/keys-DeepSeekV4Flash-Vision-EXP-ablit \
  --local-dir ~/models/keys-DeepSeekV4Flash-Vision-EXP-ablit

GPU memory utilization ≤ 0.85 .

License

Inherited from DeepSeek-V4-Flash-Vision-Exp. You must still comply with the Responsible Use gate above.