license: mit base_model:

• bot-lab-21/DeepSeek-V4.1-Flash-EXL3-3.5bpw-Pollard
• distributedcognition/DeepSeek-V4.1-Flash-abliterated base_model_relation: merge pipeline_tag: image-text-to-text library_name: vllm tags:
• deepseek_v41
• deepseek-v4.1-flash
• exl3
• exllamav3
• vllm
• mixture-of-experts
• abliteration
• uncensored
• speculative-decoding
• dspark

Important - Responsible use

This is an experimental derivative meant for research purposes only in sandboxed environments. It is not intended for use in production applications.

This is an uncensored research model whose refusal behavior was deliberately altered at the weight level. It can answer requests that the stock model may decline, including harmful or illegal requests. It has no dependable safety boundary of its own. Deployers are responsible for access controls, input and output moderation, logging, applicable law, and downstream use.

By downloading and using this model you accept all responsibility and hold harmless myself and Deepseek for any damage as a consequence, intentional or accidental.

DeepSeek V4.1 Flash Abliterated — EXL3 3.51 bpw

This is an abliterated hybrid EXL3 checkpoint of deepseek-ai/DeepSeek-V4.1-Flash. It combines:

• the mixed-K, sensitivity-allocated 3.51 bpw routed experts from bot-lab-21/DeepSeek-V4.1-Flash-EXL3-3.5bpw-Pollard;
• the scale-3.0, norm-preserving biprojected abliteration tensors from distributedcognition/DeepSeek-V4.1-Flash-abliterated;
• the original source-format non-routed weights, vision tensors, DSpark/MTP drafter, and FP8 Engram tables.

The 3.51 bpw figure applies to the routed-expert projection weights, not to every parameter or to the complete download. Sensitive and non-routed tensors retain their source representations.

What was changed

1. Routed-expert quantization

All 40 trunk layers use Pollard's EXL3 trellis representation for routed expert w1, w3, and w2 projections. The allocation is:

Attention, routers, embeddings, shared/non-routed components, vision, MTP, and Engram were not converted into ordinary EXL3 routed-expert tensors.

2. Abliteration graft

Exactly 132 source tensors were preserved from the distributedcognition checkpoint: four tensors in each layer 4–36 inclusive.

• layers.N.attn.wo_b.weight
• layers.N.attn.wo_b.scale
• layers.N.ffn.shared_experts.w2.weight
• layers.N.ffn.shared_experts.w2.scale

These weight/scale pairs remain in the donor's FP8 E4M3 plus E8M0-scale representation. The graft payload is 1,775,136,000 bytes. The routed experts are byte-identical to the Pollard substrate and were not requantized.

The intervention used norm-preserving biprojected abliteration at scale 3.0. See abliteration_config.json

3. Engram and DSpark

• Shards 47–48 contain the two original FP8 Engram lookup tables. They are unquantized and total 203,073,077,576 bytes (189.13 GiB).
• Draft layers 40–42 remain in the original source MXFP4 representation.
• The validated runtime used DSpark with five speculative tokens, probabilistic drafting, block rejection, and adaptive verification disabled.

Download size and layout

The repository is intentionally complete: download all 48 shards for the published index to resolve. A runtime may keep the Engram tables file-backed, but the files are still part of the checkpoint.

hf download UltimateIntent/DeepSeek-V4.1-Flash-Abliterated-EXL3-3.51bpw \
  --local-dir ./DeepSeek-V4.1-Flash-abliterated-EXL3-3.51bpw

Local qualification results

The same exact items were used for the capability comparison against neko-legends/GLM-5.3-Flash-Uncensored-EXL3 as baseline.

Important qualifications:

• MMLU was a deterministic, square-root-stratified 1,000-item sample across all 57 subjects, seed 47. The exact-item paired delta is the primary local comparison. It is not the same estimand as the published full-test micro-average.
• MMLU used first-token A/B/C/D logprobs with reasoning disabled.
• GSM8K used zero-shot greedy chat with maximum reasoning and no explicit completion-token cap; two of 200 responses exhausted the 8K context.
• Refusal and chat used DeepSeek numeric reasoning effort 25. The GLM baseline used its maximum-reasoning protocol, so those behavioral legs are not a formal paired cross-model comparison.
• The refusal suite is a local 100-prompt gradient and the final label was a human judgment. Keyword matching over-counted 16 responses and was not used as the verdict.

The distributedcognition donor separately reports 88.95% on the complete 14,042-item MMLU test set for both base and abliterated weights. That published donor result is not relabeled here as a measurement of this EXL3 checkpoint.

Runtime compatibility

This is a weights repository, not a self-contained inference stack. Stock Transformers and stock vLLM do not understand this checkpoint's full DeepSeek V4.1 + EXL3 + disk-Engram path.

The validated local stack used:

• vLLM revision 988d9b6777d077f843cd2164a222ac8535d36ed2 plus the qualified DeepSeek V4.1 Engram/UVA/DSpark/SM120 overlays;
• vcruz305/vllm-exl3 revision 814d4fe38082cddd838b45418c7d13a95395a36a plus the qualified V4.1 source-dispatch/UVA changes;
• turboderp-org/exllamav3 revision be57335b087e4f001c5caae061544df3c06ba01e;
• PyTorch 2.13.0+cu130, CUDA toolkit 13.4, and SM120 kernels.

The exact two-GPU qualification contract was TP2/EP2, eager execution, text-only, 8K context, one sequence, SHM NCCL transport, UVA expert layers 15,19,23,25-39, 44 GiB CPU offload allowance, disk-backed Engram, and DSpark-5. The runtime supplied these two metadata-only draft-dispatch fields at launch while preserving the full existing quantization dictionary:

{
  "mtp_experts": "source",
  "mtp_experts_start_layer": 40
}

Do not replace quantization_config with only those two fields. They are an addition to the full dictionary already present in config.json and quantization_config.json.

Pollard's model card and the WestWaters/pollard-weights project are the best public references for the EXL3 routed-expert method and the original V4.1 serving work. Runtime support is evolving quickly; verify the required architecture, source-MXFP4 draft dispatch, Engram behavior, and CUDA target before attempting a load.

Known limitations

• The local production qualification was text-only. Vision tensors are present and unchanged, but this hybrid's vision path was not requalified.
• Local serving was qualified at 8K and one sequence. The upstream architecture supports much longer contexts, but long-context behavior was not claimed for this two-GPU placement.
• Throughput depends strongly on DSpark acceptance, prompt type, PCIe topology, and Engram storage latency.
• TP2/EP2 output nondeterminism was observed even with greedy external sampling; it was transport-independent in the qualification work.
• This model can produce incorrect, harmful, biased, or fabricated content.

Credits

• DeepSeek-AI for DeepSeek V4.1 Flash.
• distributedcognition for the scale-3.0 abliteration work and evaluation.
• bot-lab-21 for the 3.51-bpw EXL3 substrate and ledgers.
• WestWaters/pollard-weights for the Pollard methodology.
• vcruz305/vllm-exl3 and turboderp-org/exllamav3 for EXL3 runtime support.
• neko-legends for the GLM5.3-Flash Uncensored model used for comparison.

License

MIT, following the upstream checkpoint. See LICENSE.