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-EXL3-2.9bpw-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)
The dealignai uncensored 1017 bundle, made loadable on the TensorFold Zig TP4 engine
running DeepSeek-V4.1-Flash across four NVIDIA DGX Spark (GB10) nodes — the exact stack Neko Legends serves.
Base: LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw →
uncensored by dealignai/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED →
this pack. Upstream model: deepseek-ai/DeepSeek-V4.1-Flash-1017.
Why this derivative exists
The TensorFold Zig TP4 engine cannot load the dealignai pack as shipped, for two reasons:
- No FP8 in the engine's DType parser. The two ~100 GB Engram tables ship as
F8_E4M3. They are dropped here;
this stack serves Engram from a separate pack (dsv41-1017-engram-w4). - The engine requires EXL3 trellis for every linear. The 54 abliterated matrices ship as dense BF16 with no
.trellis(the trellis versions were removed during abliteration).
So 54 modules — layers.10..36.attn.wo_b and layers.10..36.ffn.shared_experts.w2 — were re-encoded from their
dense BF16 form to EXL3 trellis K=5 mul1, reusing the base pack's suh/svh/mul1 scales.
Everything else is byte-identical to the base pack. Verified: name set identical (192,495 tensors), dtypes and
shapes identical, 43 shards clean, no F8 tensors, 54/54 trellis shapes match, all non-edited tensors byte-identical.
Edit retention ⟨dequant(new) − dequant(base), Δ⟩ / ‖Δ‖² with Δ = W_uncens − dequant(base):
min 0.9635 · mean 0.9996 · max 1.0209 across the 54 modules.
Benchmarks
Same hardware the model is published for: 4× DGX Spark (GB10), TensorFold Zig TP4 over RoCE, 524,288 context,
4 slots. Capability numbers use the same 500-question MMLU subset and the same 164 HumanEval problems for every
model compared, greedy decoding.
Accuracy (MMLU-500, 5-shot)
| this pack | official 1017 (local, censored) | 1017 via Venice API | |
|---|---|---|---|
| raw completions harness | 81.2 % | 86.2 % | not available |
| chat harness | 79.6 % | — | 86.6 % |
| STEM | 80.7 / 77.2 | 83.6 | 83.6 |
| Humanities | 71.7 / 68.1 | 82.3 | 85.0 |
| Social Sciences | 89.3 / 87.4 | 94.2 | 92.2 |
| Other | 84.1 / 87.6 | 86.7 | 87.6 |
moral_scenarios (n=22) |
40.9 % | 95.5 % | 95.5 % |
(each local cell shows raw-completions / chat harness; the hosted column is chat only — their API has no/v1/completions, so numbers from different harnesses are never mixed.)
The accuracy cost of uncensoring here is concentrated in normative judgement, not in knowledge: STEM and Other
move least, and moral_scenarios is where it lives. That is the trade you are taking this model for.
Coding (HumanEval-164, pass@1, sandboxed official tests)
| harness | this pack | official 1017 via Venice API |
|---|---|---|
| chat | 92.68 % | 90.85 % |
| raw completions | 76.83 % | not available |
Uncensoring did not cost code synthesis — on the identical harness this pack scores above the hosted official
model. (The raw-completions number is lower because a raw transport drops the prompt's in-body anchor and bodies
arrive unindented; that is a prompt-transport artifact, documented in our bench repo, not a capability difference.)
Speed (single stream, 4× Spark TP4)
| tok/s | |
|---|---|
| decode, short chat / 2k code / 8k prose (effort low) | 69.3 / 127.7 / 63.9 |
| decode, same, effort high | 71.7 / 91.0 / 63.0 |
| prefill, 5,296-token prompt | 2,908 tok/s (TTFT 1.82 s) |
| prefill, 21,058-token prompt | 5,033 tok/s (TTFT 4.18 s) |
Aggregate throughput with all four slots decoding is roughly 4× the single-stream figure.
Running it
# 4 ranks, one per Spark; ENGRAM_DIR is the separate Engram pack, ASSETS_DIR is the 512K rope/aot set for the rank
MODEL_DIR=/models/neko-legends-…-2.9bpw-UNCENSORED-4xSpark \
ENGRAM_DIR=/engram/dsv41-1017-engram-w4 \
ASSETS_DIR=/assets/rank-N \
TF_DSV41_CONTEXT=524288 TF_DSV41_PREFILL_PIPE=1024 \
tf-dsv41-m1 follow /model /assets # engine: github.com/neko-legends/tensorfold-4x-sparks (branch zig-tp4)
Prepared images are written on first boot (~55.1 GiB per rank) and reused afterwards. The prefill pipeline needs the
stage copies resident too (~25 GiB); on 4 Sparks this pack fits both, which is one of the reasons it is the
recommended runtime over our 3.5 bpw sibling.
A sibling exists
neko-legends/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-UNCENSORED-4xSpark re-encodes the same 54 matrices at K=6.
It recovers a few points of moral_scenarios (40.9 → 45.5 %) but decodes ~10 % slower and cannot afford the prefill
pipeline on 4 Sparks (prepared + stages exceed memory), costing 76 % of 21k prefill. We run the 2.9.
Credits
- Uncensoring:
dealignai— their abliterated 1017 bundle is the substance here. - 2.9 bpw EXL3 pack:
LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw. - 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. It is published for research, local inference and
work on uncensored assistants. You are responsible for how you use it, and nothing here is legal advice. If you
need a model that declines harmful requests, use the official DeepSeek release instead.