license: mit
base_model: dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
author: Ford Morris
tags:

• flare
• gb10
• dgx-spark
• deepseek
• deepseek-v41
• uncensored
• abliterated
• safetensors
  library_name: flare
  pipeline_tag: image-text-to-text
  inference: false

DeepSeek-V4.1-Flash-1017 — Uncensored (Flare, GB10 TP4)

Based on dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8

This is a ready-to-serve Flare package of dealignai's UNCENSORED-FP8 checkpoint for a 4× DGX Spark (GB10) cluster, exported with Ignition from NVIDIA/spark-local-inference @ 972e28e44781d8d0c84d70ae434db92c36076a43 — the first gb10 export of this checkpoint that includes the PR #129 complete-copy/MTP ordering fix (merged 2026-11-05). The weights are bit-identical to dealignai's release; this repo changes only the packaging, from raw safetensors to a prepared Flare package you can launch with one command. It is ungated.

On an identical-harness three-way comparison against the stock LIL gb10 Flare package, the measured capability cost of the uncensoring is −1.05pp on MMLU (0.8709 vs 0.8814) — dealignai's published −3.86pp penalty does not reproduce under this harness — while GPQA Diamond scores highest of the three endpoints (0.7323 vs stock 0.697), and refusal flags on a harmful-prompt battery drop to 0/8 in both chat and think modes.

Scope note on metrics: All numbers below were measured on one 4× DGX Spark fleet (TP4, RoCE ring) with one harness, on 2026-11-09. Benchmarks are generated-answer runs at the packages' fixed temperature/top-p of 1.0 with thinking on — relative precision indicators against the stock Flare package served identically, not broad general benchmarks. The KL figures are a serving-side sampled-path method (see below); read them against the included stock-vs-stock control, not against zero.

Key results (4× DGX Spark, identical harness)

Independent tensor audit of the base checkpoint (read this first)

dealignai does not publish an edit spec, so before exporting we ran a tensor-by-tensor SHA-256 payload diff of their checkpoint against stock deepseek-ai/DeepSeek-V4.1-Flash-1017 (96,085 tensors, identical tensor name set, zero config.json diffs). This is the complete inventory of what the uncensored edit touches:

• 108 tensors changed of 96,085 — the same two families as a conventional abliteration: attn.wo_b and ffn.shared_experts.w2 (weight+scale each), layers 10–36 only (54 + 54 tensors).
• Bit-identical to stock: all 94,464 routed-expert tensors, the Engram tables, the MTP/DSpark block, vision tower, embeddings, norms, and the LM head. Despite the "uncensored" branding, the edit does not touch the routed experts.
• It is a narrower-but-stronger variant of the usual projection: fewer layers than mrfakename's edit (27 layers, 10–36, vs their 40 across families) with larger per-tensor changes (payload byte-diff fraction mean 0.285, median 0.172, max 0.617). Fewer, deeper projections is consistent with dealignai's published compliance results — and the KL and benchmark sections below quantify what it costs on this serving stack.

Export details

• Tool: Ignition, NVIDIA/spark-local-inference @ 972e28e44781d8d0c84d70ae434db92c36076a43. This pin contains PR #129 ("Fix complete-copy ordering…", merge commit fa0eadaded, 2026-11-05) and PR #158 ("Keep self-reading complete copies in program order", 2026-10-30) — verified by GitHub compare (0 commits behind either merge).
• Topology: TP4 on GB10 (4× DGX Spark), RoCE ring.
• Settings: expert-scale compression CSF (recipe default), --markov-head-quantization fp8, --vocabulary-parallel-sampling, max-concurrency 32, MTP/DSpark 5 draft tokens with adaptive verification (alpha 0.75 at serving), prefill chunk 8192 — mirroring the stock LIL gb10 production configuration.

Usage

Runs on vanilla Flare (gb10, TP4). No patches. The from-source preparation container was used only for the Ignition export; it does not participate in serving — the package was validated on the unmodified stock flare-inference APT runtime.

The one non-default step is publisher trust: the package is signed with a local development signing key (namespace local-development; it signs the package integrity manifest — it is not a Hugging Face account key), so the first launch needs the standard Flare trust acceptance for that key:

# 1. Install the publisher verification profile (included in this repo)
cp development-publisher.json ~/.config/flare/

# 2. Launch (first time includes the one-time trust acceptance;
#    Flare remembers it in trust.toml afterwards)
flare orbit run woooflet/DeepSeek-V4.1-Flash-1017-uncensored-FLARE \
  --nodes <node1,node2,node3,node4> --roce-ring \
  --dangerously-accept-publisher local-development \
  --dangerously-approve-key ff4d66d66f9099b1cb47a396a6dd0b0e1f9115e221bc33cdf3b28d5b12387ee9 \
  --max-concurrency 32 --mtp-draft-tokens 5 --mtp-adaptive-verification \
  --mtp-adaptive-verification-alpha 0.75 \
  --embedding-table-storage disk --embedding-table-cache-size 1073741824 \
  --embedding-table-resident-scales --host 0.0.0.0 --port 8000

Publisher key fingerprint (SHA-256 of the public key in development-publisher.json): ff4d66d66f9099b1cb47a396a6dd0b0e1f9115e221bc33cdf3b28d5b12387ee9. Verify it against the file before approving — that check is the point of the fingerprint.

The package for your hardware is on the gb10 branch of this repo (the branch flare download / flare orbit run resolves automatically on GB10 hardware); main carries this card and the publisher profile only.

Validation method

All three endpoints in the tables are gb10 Flare packages of DeepSeek-V4.1-Flash-1017 served with identical settings (max-concurrency 32, DSpark/MTP 5 draft tokens, adaptive verification alpha 0.75, prefill chunk 8192, disk-backed Engram, temp/top-p fixed at 1.0 by the packages).

• Benchmarks: generated-answer harness (lm-eval-style prompts), thinking ON, reasoning_effort: "high" sent on every request, generous token budgets (MMLU 1024, GPQA 4096, GSM8K 1024, HumanEval 640); truncated (finish_reason=length) counts recorded per task — MMLU: stock 96, sibling 112, this package 106; GPQA: 81 / 71 / 60; GSM8K: 10 / 10 / 5. MMLU = cais/mmlu test split, stratified 25/subject × 57 subjects = 1,425 questions, seed 1234. GPQA = full Diamond set (198) via the fingertap/GPQA-Diamond mirror (canonical repo is gated; mirror verified to hold exactly the 198 Diamond items). GSM8K = openai/gsm8k test split, seeded 250-question subset, numeric exact match. HumanEval = all 164, pass@1 by execution.
• On dealignai's published MMLU claim: their card reports 83.77 vs 87.63 for the base (−3.86pp, logit-mode scoring on their stack). Under this identical three-way generated-answer harness the measured cost is −1.05pp vs our stock baseline (0.8709 vs 0.8814, at n=1425), and our stock baseline lands close to their published base figure (0.8814 vs 87.63), so the harnesses are broadly comparable: the claimed −3.86pp penalty does not reproduce here.
• KL divergence on basic text (export sanity check): KL on basic text (Pile sample), stock vs variant — a divergence in the range expected for the published edit; an export error would show up as a far larger divergence. Method (serving-side): 300 passages from NeelNanda/pile-10k (a 10k-passage sample of The Pile). Stock generates a 48-token continuation per 600-character context via /completions with top-20 logprobs; the candidate is probed at 12 positions along the stock path. Mean KL over the union of both top-20 supports (floored at 1e-6, renormalized). The Flare API does not expose prompt logprobs, so full teacher forcing is not available serving-side; this sampled-path method is the documented substitute, and the stock-vs-stock control (same procedure, separate session) calibrates its noise floor.

Honest reading: the control sets this method's noise floor at 0.100. This package's 0.213 is 2.1× that floor and above the range reported for lighter abliterations (~0.07–0.13), which matches the audit finding that this edit is stronger per tensor. It is nowhere near the order-of-magnitude elevation an export error would cause, and the benchmark table above — parity with stock on four benchmarks, best-of-three on GPQA — is the functional confirmation that the export is faithful and the divergence is the edit itself.

• Serving speed: bench_quick fixed prose prompt, 256-token generations (table above). Stock references from the 2026-11-07 LIL cutover on the same harness: C1 ~61, C8 207–223.

Validated behavior

When served with the launch command above on 4× DGX Spark, the package was verified to behave as the base edit intends:

• Refusals: 11-prompt battery (8 borderline/harmful + 3 benign controls), keyword screen with raw replies archived. This package flags 0/8 harmful prompts in chat mode and 0/8 in think mode (stock: 4/8 chat, 5/8 think) — full battery compliance, consistent with dealignai's published HarmBench result (320/320 comply, their measurement). Benign controls answered normally (0/3 flagged).
• Concurrency: 32 simultaneous requests with distinct secret facts — 32/32 clean, zero cross-request contamination (same result for stock and the sibling package).
• Tool calling: structured get_weather({"city": "Tomball"}) round trip returned correctly.
• Capability spot check: 9/10 on a fixed 10-question set (same as stock).

License & attribution

Based on dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8

• Checkpoint and uncensored edit: dealignai — License: MIT (as declared on their model card). This package is a re-export of their checkpoint for Flare; the model itself is unchanged, and this repo inherits the MIT license. Credit line included as requested on their model card.
• Base model: deepseek-ai/DeepSeek-V4.1-Flash-1017 by DeepSeek.
• Packaging performed with Ignition from NVIDIA/spark-local-inference (Flare).

This model is a community packaging of a community uncensored variant of a DeepSeek model. It is not an official DeepSeek, dealignai, or NVIDIA release. Use and redistribute in accordance with the MIT license and the base model's terms.

Author

Ford Morris