DeepSeek-V4.1-Flash-Abliterated-GGUF

DwarfStar (ds4) GGUFs of community-abliterated DeepSeek V4.1 Flash checkpoints. They are not general GGUFs: the tensor layout, quantization mix and Engram packing are ds4-specific.

All four files are 365 713 686 528 bytes. The two Engram tables (about 189 GiB) are packed at the end of each file and read from disk as needed, never resident, so a fast local SSD matters.

Maya-XS: DeepSeek V4.1 Flash 1017, abliterated

File: DeepSeek-V4.1-Flash-1017-Maya-XS-Abliterated.gguf. SHA-256: f414995b3797f77009634fac995f50834e70393771f9423fca984c892503b42a.

This native ds4 conversion uses Project Maya's quantization at revision d1fc01009f21e26157acb6858c315d0365483b72. It copies 120 IQ1_S expert tensors and 339 Q8 dense tensors without requantization. It retains the native 1017 metadata, small unquantized tensors and disk-only Engram tables. The model keeps about 107.35 GiB of weights resident on Metal.

The existing pqhaz Heretic edit replaces 38 attention output tensors in blocks 2-39. Every source and replacement hash was checked. Readback verified all 38 writes and a byte comparison found the other 105.78 GiB of resident data unchanged. English, Traditional Chinese and retrieval smoke tests passed on the abliterated file with no swap growth.

Paired quality and speed measurements

The quality comparison below uses the unablated Maya conversion and the old unablated Q1, on 100 September V4.1 continuation fixtures (6160 tokens). The fixtures support a relative comparison, not a measurement of fidelity to the original 1017 weights. The abliterated Maya file has not run the full scoring suite.

NLL was 16.34% lower. API agreement excludes the fixture's documented tokenizer mismatch in case_047 (99 aligned cases). Speed is one back-to-back CLI sample on a 128 GB M5 Max, without vision.

Use and context limits

Requires audreyt/ds4 at 0dff39f91721bcf3472813c6d2bebf9353df2a1a or later for Q8 embeddings, compressor weights and indexer weights. Raise the Metal wired-memory ceiling as described below. No SSD streaming or speculative draft model is used.

./ds4 -m DeepSeek-V4.1-Flash-1017-Maya-XS-Abliterated.gguf --ctx 262144
./ds4-server -m DeepSeek-V4.1-Flash-1017-Maya-XS-Abliterated.gguf --ctx 128000 \
  --vision DeepSeek-V4.1-Flash-1017-Vision.gguf

On a 128 GB M5 Max, text-only context allocation 262144 passed with 512-row prefill. A 10237-token prompt retrieved its verification code correctly. This was not a full-262k-depth test. Allocation 393216 was refused by the 112 GiB guard. With the vision sidecar and candor steering, allocation 131072 passed; pi-ds4 v0.7.3 uses 128000 by default.

Q1: DeepSeek V4.1 Flash 1017, abliterated

DeepSeek-V4.1-Flash-Abliterated-Q1.gguf, SHA-256 d302c2caf7f12766eba2c431fc852261ddd0bf4d30b2760697d6036dac0f3cbd.

The abliteration is pqhaz's Heretic edit of deepseek-ai/DeepSeek-V4.1-Flash-1017: 38 tensors, the attention output projection attn.wo_b in blocks 2–39, refusal direction removed at strength 2.5. pqhaz reports held-out refusals with reasoning on going from 92/100 to 1/100, with KL 0.136 on harmless prompts, and no capability benchmarks.

Recipe

Built with gguf-tools/mixed/deepseek41_q1_from_unsloth.py (audreyt/ds4) as an APFS clone of the ds4-native 1017 Q2 (audreyt/deepseek-v4.1-flash-1017-gguf), with tensors replaced in place:

• Routed MoE experts: IQ1_S from unsloth/DeepSeek-V4.1-Flash-1017-GGUF UD-IQ1_S (revision 60ad365). 88 expert tensors are used as published; the 32 that Unsloth ships as IQ2_XXS are requantized to IQ1_S with Unsloth's imatrix so the file fits ds4's 112 GiB V4.1 budget. That step requantizes already-quantized weights rather than starting from the original checkpoint.
• Dense weights: Unsloth's own tensors (mostly Q6_K/Q5_K) for attention q_a/q_b/kv, shared-expert gate/up/down, output head, token embedding and the Engram projection.
• Attention output attn_output_b: Q8_0, the 38 abliterated tensors from the pqhaz checkpoint.
• Everything else, including the Engram tables, comes from the ds4-native 1017 Q2 unchanged. The metadata keeps the 1017 source revision, so the 1017 vision sidecar pairs with this file.

Use

Needs audreyt/ds4 at d3f4cd6 or later (IQ1_S and Q5_K/Q6_K dense weights run only on Metal; CUDA and ROCm refuse the file). On a 128 GB Mac, first raise Metal's wired-memory ceiling: macOS leaves about 107.5 GiB by default, less than this file plus the vision sidecar need under ds4's 112 GiB budget.

sudo sysctl iogpu.wired_limit_mb=114688   # add to /etc/sysctl.conf to persist

For images, add DeepSeek-V4.1-Flash-1017-Vision.gguf from audreyt/deepseek-v4.1-flash-1017-gguf.

./ds4 -m DeepSeek-V4.1-Flash-Abliterated-Q1.gguf --ctx 8192
./ds4-server -m DeepSeek-V4.1-Flash-Abliterated-Q1.gguf --ctx 256000 \
  --vision DeepSeek-V4.1-Flash-1017-Vision.gguf

pi-ds4 installs and runs this pair on a 128 GB Mac.

On a 128 GB M5 Max this file decodes at about 37.9 t/s (ds4 CLI, ctx 8192, 128 greedy tokens, High Power energy mode; Low Power is far slower). Without the vision sidecar, the 256 k window opens with a 2048-row prefill batch, 384 k with 1024 rows, and 512 k or more does not fit next to the weights. The sidecar takes another 0.9 GiB: with vision the 256 k window opens with 1024-row prefill chunks (measured), and the ceiling is lower. With the sidecar at that window, a 533-token image turn prefilled at 194 t/s and read a strict OCR fixture back exactly.

No official-continuation quality score has been run for this Q1 yet. It is a compromise that fits one 128 GB Mac, not a quality upgrade over the Q2 files.

Q2: DeepSeek V4.1 Flash, abliterated

DwarfStar Q2 GGUF of the community-abliterated DeepSeek V4.1 Flash checkpoint (s-zaizen/DeepSeek-V4.1-Flash-Abliterated, itself abliterated from deepseek-ai/DeepSeek-V4.1-Flash with Heretic; refusal keyword rate reported there as 97/100 → 24/100).

Two files: ...-imatrix.gguf (recommended) and the original DeepSeek-V4.1-Flash-Abliterated-Q2.gguf bootstrap build (kept for reference; the imatrix build supersedes it).

Recipe

Converted with deepseek41_quantize.py (DwarfStar ds4 repo) using the project's C quantizers, --quant q2, imatrix-calibrated routed experts (16-prompt / 26K-token calibration collected on the abliterated Q2):

• Routed MoE experts: IQ2_XXS gate/up, Q2_K down
• Attention, shared experts, projections, output: Q8_0 / F16
• Engram FP8 rows packed losslessly at end of file (read from disk as needed, never resident)

deepseek41_validate_gguf.py layout audit: PASS (1046 tensors).

Quality vs stock Q2

Target-token NLL on 100 tracked official V4.1 Flash continuations (deepseek-v4.1-flash-20260910-general, ctx 4096, SSD streaming), same harness for both files:

Essentially at parity with stock: the remaining ~0.01 NLL is plausibly the ablation itself. First-token match and top-1 agreement beat stock within noise.

Use

./download_model.sh ds41f-q2  # stock weights instead (341 GiB)
./ds4 -m DeepSeek-V4.1-Flash-Abliterated-Q2-imatrix.gguf --ssd-streaming --ctx 32768

On one 128 GB Mac, use --ssd-streaming (152 GiB main weights; Engram tables stay on disk). Needs a fast local SSD.

License

Base weights © DeepSeek, MIT (retained via the source checkpoints, both MIT). Maya-derived tensors by Peasant Smith and Unsloth's quantized tensors are used under the same license. Conversion tooling per the ds4 repo license.