DeepSeek V4 Flash Vision-Exp Abliterated (ds4 IQ2)

Official-recipe Vision-Exp language GGUF (IQ2_XXS routed experts, Q2_K down, Q8_0 attention/shared/output) with the published rank-1 refusal-direction edit baked into 33 blk.{10..42}.attn_output_b.weight tensors.

This is not Headroom128 0731 Flash. Do not attach the 0731 DSpark support GGUF. Pair with the unmodified 316-tensor encoder.

Files

Language KV: 1328 tensors / 66 KV, architecture=deepseek4, checkpoint_variant=vision-exp, sidecar_required=true, source.revision=e46e16bf6035c6f317eb2ac7458eb0362926d402.

How it was made

1. APFS-clone the official antirez Vision-Exp IQ2 language GGUF.
2. HTTP-Range the 33 already-quantized Q8_0 attn_output_b payloads from apetersson/DeepSeek-V4-Flash-Vision-Exp-Abliterated revision f41380effb3b78f9e08cec2110296a149fb2e063 (Basic128-Routed-IQ2_M), Content-Range total 102826285248.
3. pwrite each payload at the parsed local data_pos+rel_offset (not a 35651584-byte stride).
4. SHA-256 every destination payload against the pinned Basic128 source. Control blk.0.attn_output_b.weight unchanged (384e5056…).

Abliteration recipe (publisher): rank-1 residual on HF layers.N.attn.wo_b for N=10..42, λ=3.5, preserve output-row L2; MTP/shared-down untouched. Direction transferred from drowzeys 0731 DSpark ablit 85ae5bba.

Run (ds4 origin/main / audreyt d0c2b43)

./ds4 -m DeepSeek-V4-Flash-Vision-Exp-Abliterated-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8.gguf \
  --vision DeepSeek-V4-Flash-Vision-Encoder.gguf \
  --nothink --temp 0
# then: /read photo.jpg

M5 Max 128 GiB smoke (2026-09-01): image 512x507, 169 image tokens, Earth/Africa globe caption, 285.75 prefill / 44.94 gen t/s.

License

MIT, inheriting DeepSeek V4 Flash / Vision-Exp. Derivative of deepseek-ai/DeepSeek-V4-Flash-Vision-Exp and the official antirez IQ2 quant. Abliterated wo_b payloads from apetersson Basic128 (same MIT inherit).