license: mit
base_model:
- TensorFold/DeepSeek-V4-Flash-0731-MXFP4-MLX
- prem-research/DeepSeek-V4-Flash-0731-abliterated
library_name: mlx
tags: - mlx
- abliterated
- deepseek_v4
- mxfp4
pipeline_tag: text-generation
DeepSeek-V4-Flash-0731 MXFP4-MLX — abliterated
This is prem-research/DeepSeek-V4-Flash-0731-abliterated
(revision 457210de) in the MLX layout of
TensorFold/DeepSeek-V4-Flash-0731-MXFP4-MLX (revision d6be051a).
Both derive from deepseek-ai/DeepSeek-V4-Flash-0731.
How it was made
TensorFold/DeepSeek-V4-Flash-0731-MXFP4-MLX is a byte transplant of deepseek-ai/DeepSeek-V4-Flash-0731: fp4 e2m1 expert weights stored as MLX
mxfp4 with the same bytes, fp8 e4m3 dense weights as MLX mxfp8 with the same bytes and each 128x128
block scale repeated per 32-group, every other tensor unchanged under the same name.
prem-research/DeepSeek-V4-Flash-0731-abliterated keeps the original format and rewrites only the shards of layers 17-36.
Method: ARA (prem-research model card): attn.wo_b, ffn.shared_experts.w2 and the listed experts' ffn.experts.E.w2, layers 17-36.
This repository is the base with the tensors that changed replaced by the same transplant:
- every tensor of the rewritten source shards was transplanted and compared with the base:
320 of 31,450 differ, all inside the steered modules; the other
31,130 are byte-identical to the base, which re-proves the transplant on this data; - the 320 changed tensors were written in place into 15 base shards;
headers, offsets, shapes and dtypes did not move; - every other tensor and file is byte-identical to TensorFold/DeepSeek-V4-Flash-0731-MXFP4-MLX at
d6be051a,
except this README; - numeric check: each changed module dequantized with
mx.dequantizefrom this layout against the
fp8/fp4 source, max abs error 0.0.
Recipe: byte transplant: fp4 I8->U32 and fp8 F8_E4M3->U32 same bytes; fp4 scales same bytes; fp8 128x128 block scales repeated to [R, C/32]; other tensors unchanged.
The splice is scripts/lab/dsv4/dsv4_abliterated_splice.py, run byscripts/lab/ablit-perf/job-dsv4-abliterated.sh in the MacProvider repository.
Checks this artifact passed
The publish step runs only after the job passed all of its gates:
- token parity of the MacProvider native engine against DeepSeek's reference implementation on these
weights at 1K and 8K context: first 16 top-1 tokens exact and at least 240 of 256 greedy tokens equal; - refusal sanity (greedy, keyword classes): the base refused 20 of 20 harmful prompts;
this model refused 0 of 20 harmful and 0 of 20 harmless prompts.
These are lab sanity gates, not a capability or safety evaluation.
License
MIT, the license of both TensorFold/DeepSeek-V4-Flash-0731-MXFP4-MLX and prem-research/DeepSeek-V4-Flash-0731-abliterated (and of deepseek-ai/DeepSeek-V4-Flash-0731).