language:
- en
- es
- fr
- ru
- it
- th
license: other
license_name: deepseek-v41-flash-1017
license_link: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/LICENSE
base_model: dealignai/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-UNCENSORED
base_model_relation: quantized
library_name: exllamav3
pipeline_tag: text-generation
tags: - exllamav3
- exl3
- deepseek
- deepseek-v41
- uncensored
- abliterated
- tensorfold
- dgx-spark
- gb10
- 4x-spark
DeepSeek-V4.1-Flash-1017 — UNCENSORED (Neko Legends 4× DGX Spark edition, 3.5 bpw)
The higher-fidelity sibling ofneko-legends/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED-4xSpark.
Same uncensoring, more bits on the 54 abliterated matrices.
Base: LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw →
uncensored by dealignai/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-UNCENSORED →
this pack. Upstream model: deepseek-ai/DeepSeek-V4.1-Flash-1017.
Read this before choosing it: on four Sparks we run the 2.9 bpw pack, not this one. This pack is a fidelity
variant, and the reason is in the speed table below.
Why this derivative exists
The TensorFold Zig TP4 engine cannot load the dealignai pack as shipped:
- The two ~100 GB Engram tables ship as
F8_E4M3, which the engine's DType parser does not support. They are dropped
here; this stack serves Engram from a separate pack. - The 54 abliterated matrices (
layers.10..36.attn.wo_b,layers.10..36.ffn.shared_experts.w2) ship as dense BF16
with no.trellis, and the engine requires trellis for every linear.
Those 54 modules were re-encoded to EXL3 trellis K=6 mul1 (the 2.9 pack uses K=5). Everything else is
byte-identical to the 3.5 bpw base pack.
Edit retention: min 0.9977 · mean 0.9992 · max 0.9998. Quantization error on the edited matrices is roughly 40%
lower than the K=5 encode in the 2.9 pack.
Benchmarks — measured on the same 4× DGX Spark stack, same harnesses as the 2.9 pack
Accuracy and coding
| this pack (3.5 bpw) | 2.9 bpw pack | official 1017 via Venice API | |
|---|---|---|---|
| MMLU-500, raw completions | 81.6 % | 81.2 % | n/a |
| MMLU-500, chat harness | 81.0 % | 79.6 % | 86.6 % |
moral_scenarios (n=22), raw |
45.5 % | 40.9 % | — |
moral_scenarios (n=22), chat |
22.7 % | 22.7 % | 95.5 % |
| HumanEval-164 pass@1, chat | 93.29 % | 92.68 % | 90.85 % |
| HumanEval-164 pass@1, raw completions | 76.22 % | 76.83 % | n/a |
So the extra bits buy about +4.5 points on moral_scenarios under the raw harness and nothing measurable anywhere
else. Note both local packs collapse to answering "D" on all 22 moral_scenarios items under chat with thinking off —
that is a separate behavioural artifact, documented in our bench repo, not a property of this pack alone.
Speed — this is why 2.9 is the runtime
| this pack | 2.9 bpw pack | |
|---|---|---|
| decode, short chat / 2k code / 8k prose (effort low) | 62.2 / 116.0 / 56.5 tok/s | 69.3 / 127.7 / 63.9 |
| prefill, 21,058 tokens | 2,780 tok/s | 5,033 tok/s |
| prefill, 5,296 tokens | 2,426 tok/s | 2,908 tok/s |
The gap is structural, not tuning: prefill pipeline stages have to stay resident alongside the prepared images, and on
four Sparks this pack needs 67.3 GiB per rank prepared (2.9: 55.1) plus ~33 GiB of stage copies (2.9: ~25). Both do not
fit, so this pack runs with the prefill pipeline off — which costs 76% of 21k-token prefill throughput against the
2.9 pack's live configuration. With the pipeline off on both, the two packs are within 2.8% of each other.
If you have more memory per rank than a stock 4-Spark setup, or you don't care about long-prompt prefill, this pack is
the better-quality option. On stock four Sparks, we run the 2.9.
Running it
MODEL_DIR=/models/neko-legends-…-3.5bpw-UNCENSORED-4xSpark \
ENGRAM_DIR=/engram/dsv41-1017-engram-w4 \
ASSETS_DIR=/assets/rank-N \
TF_DSV41_CONTEXT=524288 \
tf-dsv41-m1 follow /model /assets # engine: github.com/neko-legends/tensorfold-4x-sparks (branch zig-tp4)
Decode/prefill are unaffected by context length settings; the pipeline knob (TF_DSV41_PREFILL_PIPE) is the one this
pack cannot afford on four Sparks.
Credits
- Uncensoring:
dealignai. - 3.5 bpw EXL3 pack:
LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw. - Base model:
deepseek-ai/DeepSeek-V4.1-Flash-1017. - Four-Spark TensorFold engine:
neko-legends/tensorfold-4x-sparks, built onashhart/TensorFoldand Jay Leaton's DeepSeek V4.1 Zig port. - Benchmarks, build and verification scripts:
bench/neko-uncens-1017-20261019/in the engine repo.
Intended use and responsibility
This model has had its refusals removed. It will produce content the official model declines to produce, including
instructions for things that are illegal or harmful where you live. Published for research, local inference and work on
uncensored assistants. You are responsible for how you use it. If you need a model that declines harmful requests,
use the official DeepSeek release.