DeepSeek-V4.1-Flash-1017-abliterated — MXFP4 GGUF

This MXFP4 GGUF conversion was made for use with DS4 / DwarfStar by antirez and DS4 for the Mac Ultra by IngeniousIdiocy.

Converted from mrfakename/DeepSeek-V4.1-Flash-1017-abliterated, based on DeepSeek-V4.1-Flash-1017. The original chart below is reproduced unchanged with credit to mrfakename. It shows the source author's Q8_0 evaluation, rather than measurements of this MXFP4 conversion.

Source model and ablation

According to mrfakename, the source uses a two-dimensional refusal subspace derived from chat and thinking prompts. Layer-specific projections modify attention output and shared-expert down weights across the 40 main blocks. Edited matrices were returned to FP8 using original block scales and stochastic rounding. Routed experts, embeddings, Engram data and draft layers were left unchanged by that ablation. See the original model card for the method and source files.

Source author's results

These Q8_0 results used 100 held-out requests and a 384-token generation budget. Thinking responses were assessed from the reasoning produced within that budget. The KL measurement includes noise from batched inference. No corresponding evaluation is claimed for these MXFP4 GGUFs.

This conversion

The Linux conversion used the v41-m3ultra branch of IngeniousIdiocy/ds4-v41-m3ultra, with deepseek41_quantize.py --quant mxfp4 and deepseek41_vision.py.

• Routed experts preserve their native MXFP4 values.
• Attention/shared/output projections use Q8_0 where specified by the converter; the complete model is mixed precision.
• Engram values and scales retain their native representation for disk lookup.
• The separate vision encoder preserves its BF16 tensor bytes.
• The ablation was already applied in the source checkpoint; this conversion performs no additional training or ablation.

Source revision: 517efab9ccfef7f714dbd7d679f181a86f43ccf9. This revision was registered in the converter's checkpoint metadata table, retaining its own source URL. The runtime converter commit has not yet been recorded in this repository.

Both DS4 projects above are intended runtime targets. Per-project inference, vision and DSpark test results for these particular files have not yet been documented here. This community conversion does not imply endorsement by the upstream authors.

Release files

The main model is distributed as six binary transport parts named DeepSeek-V4.1-Flash-1017-abliterated-MXFP4.gguf.part-000 through .part-005. Each of the first five is 100,000,000,000 bytes; the last is 1,609,136,128 bytes. Join them before use. These are transport parts, not individually loadable GGUF shards.

The drafter and vision encoder are supplied as complete files. SHA256SUMS contains the checksums of all three complete GGUFs. The SHA256SUMS.parts manifest covers all six transport parts. Their concatenated bytes have been verified against the original complete model checksum in SHA256SUMS.

Download and reconstruct

Download all six main-model parts, the smaller GGUFs you intend to use, and the checksum files from this repository's Files and versions tab. Put them in a new directory, for example gguf/abliterated-release inside your chosen DS4 checkout.

On macOS, from that download directory:

shasum -a 256 -c SHA256SUMS.parts

cat DeepSeek-V4.1-Flash-1017-abliterated-MXFP4.gguf.part-00{0,1,2,3,4,5} \
  > DeepSeek-V4.1-Flash-1017-abliterated-MXFP4.gguf

shasum -a 256 -c SHA256SUMS

On Linux, use sha256sum -c instead of shasum -a 256 -c. The full manifest expects all three GGUFs; a text-only download can verify its main model's entry separately. Stop if a checksum fails. After successful reconstruction and verification, the transport parts can be removed. Retaining both the parts and reconstructed model requires roughly 1.01 TB of storage.

Run with DS4

Build your chosen DS4 checkout with make, and run from its root. For a 512 GB Apple Silicon Ultra, begin with a modest context:

./ds4 \
  -m gguf/abliterated-release/DeepSeek-V4.1-Flash-1017-abliterated-MXFP4.gguf \
  --ctx 8192 --nothink --temp 0 \
  -p "Write a Python function that checks whether a string is a palindrome."

For DSpark speculative decoding:

./ds4 \
  -m gguf/abliterated-release/DeepSeek-V4.1-Flash-1017-abliterated-MXFP4.gguf \
  --dspark \
  --mtp-model gguf/abliterated-release/DeepSeek-V4.1-Flash-1017-abliterated-DSpark-MXFP4.gguf \
  --ctx 32768

For image input on Metal:

./ds4-agent \
  -m gguf/abliterated-release/DeepSeek-V4.1-Flash-1017-abliterated-MXFP4.gguf \
  --vision gguf/abliterated-release/DeepSeek-V4.1-Flash-1017-abliterated-Vision.gguf \
  --ctx 32768

Use the vision encoder supplied here: its source revision matches the language model. Keep the model on a fast local SSD because Engram tables are read from disk during inference. Consult official DS4's model guide and the Ultra fork's hardware notes for memory sizing and runtime support. Other GGUF engines have not been validated for this layout.

Attribution and license

Credits to DeepSeek for the original checkpoint, mrfakename for the abliterated model and original chart, antirez for DS4 / DwarfStar, and IngeniousIdiocy and contributors for the Mac Ultra fork and conversion tooling.

The inherited MIT copyright and permission notice is included in LICENSE. The abliterated source is intended to reduce refusals and can produce unsafe or incorrect output; refer to its model card for limitations.