English注册

drowzeys/keys-Qwen3.8-27B-NVFP4-DFlash2-Ablit-Cybersecurity-Unlock-1M-Context-YARN-Single-DGXSpark

🤗 Hugging Face 来源text-generationapache-2.0激活 27B-other⚠ 没有 SHA-256 权重今天更新
帮助这个模型通过 Pirate Face 分发

模型卡、文件列表和校验和已在此收录。如果你持有文件并有权分享，可以提交种子，让其他人从节点下载。

获取来源文件需要 Hugging Face 批准。
为此模型做种

keys-Qwen3.8 27B NVFP4 DFlash2 Ablit+Cybersecurity Unlock 1M Context YaRN Single-DGXSpark

Best of the community, open source — compiled in one place.

This is not a new model Keys trained. It is a measured one-DGX-Spark recipe that wires together the best public pieces for Qwen3.8-27B: AEON-7’s uncensored abliteration, a uniform NVFP4 that actually fits GB10, Inco/Z Lab DFlash2 n=7, AEON’s prebuilt vLLM image, and YaRN to 1,048,576 tokens.

Go read the original repos. That is where the research, ablit trials, drafter, and Spark image live. We only put the wiring and the bake-off numbers in one folder.
Hardware one NVIDIA DGX Spark (GB10) — not the dual-Spark Flash-Next pair
Window 1,048,576 (YaRN factor 4.0 × native 262,144)
1M needle HIT at 999,714 prompt tokens
Ablit + cyber AEON uncensored weights + Keys unlock template → cyber 8/8, refusal32 32/32
Speed (16k / seqs 64) tea 135.7 agg @ c=16 · prose decode ~20 · code ~42
Intelligence model corrected a wrong harness key (Monday, not the gold Sunday)

Prebuilt image (this is the upload):

ghcr.io/drowzeys/keys-qwen38-27b-nvfp4-dflash2-ablit-cyber-unlock-1m-yarn-single-dgxspark:latest
# digest sha256:fd31bd450510bebbb6d28623eb88ff8a86a6ac6c8db2a9009f21fafe8ff1eff9

That image is a retag of AEON-7’s ghcr.io/aeon-7/aeon-vllm-ultimate:latest
upstream digest sha256:dd2018473ed88bc23b01cfc3179b5b6896a7f0f152ae06d8d274db62d330ef48
(v0.27.1+aeon.sm121a.dspark). We did not rebuild vLLM. Layers are mounted from AEON-7.

Weights are not in this Hugging Face repo. Download them from the people who made them:

hf download sakamakismile/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4 \
  --local-dir ~/models/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4
hf download incoai/Qwen3.8-27B-DFlash2 \
  --local-dir ~/models/Qwen3.8-27B-DFlash2
docker pull ghcr.io/drowzeys/keys-qwen38-27b-nvfp4-dflash2-ablit-cyber-unlock-1m-yarn-single-dgxspark:latest
# identical to:
# docker pull ghcr.io/aeon-7/aeon-vllm-ultimate:latest

Why each community repo (read them)

Full table: CREDITS.md. Short version:

We used Because it is the best open piece for this job Link
Qwen/Qwen3.8-27B Best open dense 27B: code, agents, 262k rope, vision, thinking HF
AEON-7 Ultimate Uncensored BF16 Best coherence-first ablit of that 27B (not a lobotomy). This is the abliteration. HF · GitHub
sakamakismile NVFP4 Best uniform compressed-tensors NVFP4 of the AEON master that fits one Spark with DFlash2 + 1M KV HF
AEON-7 NVFP4-MIXED Best official Spark/5090 AEON quant — still go there. We used uniform on this box because it was +33% tea vs MIXED @ c=16 HF
Inco + Z Lab DFlash2 Best lossless block-diffusion drafter for Qwen3.8-27B; AEON Spark recipe is n=7 incoai · z-lab · blog · code
AEON aeon-vllm-ultimate Best prebuilt GB10 vLLM (TRITON_ATTN, DFlash, nested YaRN, qwen3 parsers) ghcr.io/aeon-7/aeon-vllm-ultimate:latest
vLLM Best production OpenAI-compatible server this can sit on github.com/vllm-project/vllm
YaRN Best standard RoPE stretch: 262144 × 4 = 1048576 arXiv:2309.00071

Abliteration tools AEON used (read their card): abliterix, heretic, Arditi et al. 2406.11717, FernflowerAI SSM repair.

What Keys put together

  1. Nested YaRN under text_config.rope_parameters and a YaRN overlay on the DFlash2 config.json (vLLM will not stretch dflash rope past 262k by itself).
  2. Cybersecurity-unlock default system template. AEON uncensored weights alone were 7/8 cyber / 30/32 refusal on our first B5 gate; after the template: 8/8 and 32/32.
  3. MIXED-style sampling (repetition_penalty=1.05).
  4. Measured eval on spark-13b3, 2026-09-08.

GMU 0.70 on the 1M seat, 0.75 on the 16k speed seat. Never above 0.85.

One-shot (one DGX Spark)

Accept RESPONSIBLE_USE.md, then:

git clone --depth 1 https://github.com/drowzeys/keys-Qwen3.8-27B-NVFP4-DFlash2-Ablit-Cybersecurity-Unlock-1M-Context-YARN-Single-DGXSpark
cd keys-Qwen3.8-27B-NVFP4-DFlash2-Ablit-Cybersecurity-Unlock-1M-Context-YARN-Single-DGXSpark
I_AGREE=1 bash one-shot.sh

That script: prints credits → pulls the AEON-7 prebuild (Keys GHCR, fallback to public aeon-vllm-ultimate) → downloads sakamakismile NVFP4 + Inco DFlash2 → launches 1M YaRN / seqs 2 / gmu 0.70 / DFlash2 n=7 → waits for max_model_len=1048576 → smokes PONG.

I_AGREE=1 bash one-shot.sh --speed          # 16k / seqs 64 speed seat
I_AGREE=1 bash one-shot.sh --skip-download  # weights already on disk

Health: curl -s http://127.0.0.1:8000/v1/models → aeon, max_model_len 1048576.

Short-prompt decode is not killed by the 1M cap; concurrency is (seqs 2 vs 64). Do not quote 135.7 agg on the 1M boot. GMU never above 0.85.

Measured (spark-13b3, 2026-09-08)

Engine vLLM 0.27.1+aeon.sm121a.dspark. Thinking off on eval.

16k / seqs 64 / gmu 0.75 (speed seat):
Tea wall 17.9 tok/s
Prose decode ~19–21 tok/s
Code decode 42.3 tok/s
Tea c=1 / 8 / 16 24.3 / 99.1 / 135.7 agg
Cat-6 mixed @ c=16 126.3 agg

Capability: math 8/8 · coding 5/5 actual (whitespace harness miss) · intelligence 6/6 actual · truth 8/8 · agentic 4/4 · cyber 8/8 · refusal32 32/32.

The intelligence monday item: harness gold was Sunday. The model answered Monday. Yesterday Friday → today Saturday → tomorrow Sunday → day after tomorrow Monday. That is the model correcting a wrong exam key, not a miss. MIXED “passed” by matching the bad gold.

vs AEON Ultimate MIXED (2026-09-09)

Same-day bake-off: this uniform NVFP4 B5 (.4, 16k/seqs 64) vs AEON-7 NVFP4-MIXED (.1, 16k/seqs 32). Full tables: COMPARE.md.

Keys B5 AEON MIXED
STEM 50-cat (math/physics/chem/genomics/folding/compare) 39/40 39/40
Protein folding 5/5 5/5
monday actual Monday Sunday
Cyber 10 + refusal32 10/10 · 32/32 10/10 · 32/32
Essay / list / tea decode tok/s 21.5 / 19.7 / 22.2 18.3 / 17.3 / 16.0
Tea agg c=16 / c=32 147.7 / 181.8 117.4 / 152.7

Tied on STEM and unlock. B5 wins prose speed and the concurrency sweep. MIXED got the KE item (16 J); B5 dropped the ½ (32). MIXED still matches the bad Sunday gold.

1M YaRN / seqs 2 / gmu 0.70, KV 1,113,770 tokens:

Prompt tok Needle
23,546 … 279,773 HIT (280k is already past native 262,144)
999,714 HIT NX-7459C85377 (2.89 h prefill on one Spark)

Full tables: RESULTS.md. No payloads in the public pack.

⚠️ Responsible use

Uncensored + cyber unlock. You own prompts, outputs, and harm. Gated access. See RESPONSIBLE_USE.md.

Files

File What
COMPARE.md Keys B5 vs AEON Ultimate MIXED (2026-09-09 bake-off)
one-shot.sh pull image + original weights + launch + smoke
serve_b5_on_4.sh 1M YaRN live recipe
serve_b5_16k.sh 16k / seqs 64 rollback
yarn_hf_overrides.json nested YaRN hf-overrides
dflash2_config_yarn_1m.json drafter rope overlay
chat_template_b5_uncensored.jinja cyber unlock default system
Dockerfile.prebuild FROM AEON digest (no rebuild)
CREDITS.md every repo and why
raw/ measured JSON

License: Apache-2.0 inherited from Qwen. Image: AEON-7. Weights: AEON-7 / sakamakismile / Inco / Z Lab.