license: apache-2.0
base_model: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
language:

• en
• zh
• multilingual
  library_name: transformers
  pipeline_tag: text-generation
  tags:
• 27b
• early-access
• draft
• abliterated
• aeon
• aeon-7
• nvfp4
• fp8
• mixed-precision
• modelopt
• gdn
• hybrid-attention
• multimodal
• qwen
• qwen3
• qwen3.8
• reasoning
• uncensored
• unfiltered
• vision
• vllm

Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4-MIXED

Early Access for Paid Patreon Subscribers — BF16 master, serve recipes, next cuts.

I've spent a lot of months learning what abliteration actually costs. Turns out the trap most people fall into is chasing a vanity KL of zero and calling every judge-R a leftover refusal. That is how you lobotomize a 27B. I wanted the opposite: willfully compliant, coherent, useful — and still a real model when the answer gets long.

So this is the ~23G ModelOpt mixed deploy cut of AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16. Not a blunt NVFP4 dump. Not a dequant-edit-requant circus. A lattice designed so the last layers can still write code that passes unit tests, while the bulk of the MLP rides NVFP4 where Blackwell is actually hungry for it.

Early Access Draft. Cosmic / GA is a later brand. Do not treat a long-context loop as the last word on this line.

Model family (pick your seat)

Same will. Different seats. The BF16 is the source of truth. This card is what you actually serve day to day.

The lattice (why MIXED, not "just NVFP4")

I do not trust equal-bitwidth recipes on a hybrid GDN / full-attn stack. The first Mix B bake FP8'd only the last-8 down_proj and left gate/up on NVFP4. Coding cratered to 0.694 on AEON Bench unit tests — real syntax and assert fails, not a moody judge. Leftover-KL on those miss prefixes lit up layers 56–63. So the sibling bake did the honest thing: last-8 full MLP (gate + up + down) → FP8. Coding jumped to 0.833.

Calib 1024×2048. Export gate: found=168 missing=0 last8_mlp_fp8=24. Parent BF16 untouched. Post-export: grafted unmodified mtp.* (15 BF16 tensors, 0.791G) from Ultimate BF16 — ModelOpt had dropped the head.

What you get

• Uncensored will — hall monitor gone; answers instead of sermons (same lineage as the BF16 draft)
• Coding that ships — AEON Bench Coding 0.833 vs prior Mix B 0.694 (and ahead of stock Unsloth NVFP4 Coding 0.806 on the public card)
• Agentic legs — Hermes 0.918, OpenCode 0.818 on the same local follow-on (OpenClaw still the soft spot)
• Deployable size — ~23.8G (NVFP4 body + BF16 vision + BF16 MTP) on one Spark / one 5090 / one PRO 6000
• Full multimodal + MTP — vision tower and video preprocessor intact; native MTP head for RTX speculative decode
• Honest Early Access — Intelligence (Math/Reasoning/Prose) still trails stock Qwen's 84.5 composite; Cosmic is the heal track, not this card pretending otherwise

Local AEON Suite v4 run ceaac9e418 (GB10, think 32k / max 65k / retry 131k, c=8, mothership fail-closed):

QuickStart — three machines, two images

Use ghcr.io/aeon-7/aeon-vllm-ultimate:latest on DGX Spark / GB10.
Use ghcr.io/aeon-7/aeon-vllm-ultimate-rtx:latest on RTX 5090 and RTX PRO 6000 (sm_120). Do not cross those images. Spark is sm_121a / aarch64 UMA. The RTX image is sm_120 / amd64 dedicated VRAM.

Gen defaults for all three (Qwen3.8 card): temperature 0.6, top_p 0.95, top_k 20, repetition_penalty 1.05. Thinking on: chat_template_kwargs={"enable_thinking": true, "reasoning_effort": "medium"}.

ModelOpt note: leave --quantization unset so hf_quant_config.json selects modelopt_mixed. On current Spark tags you still need the two-line #54367-style modelopt.py bind-mount (GDN in_proj_qkvz + FP8_PER_CHANNEL_PER_TOKEN) until it lands in :latest. Native MTP is included (BF16 head, 15 tensors). Do not run MTP and DFlash at the same time — pick one. Spark often still prefers external DFlash2 n=7 for throughput; RTX seats should use in-checkpoint MTP n=3.

1) DGX Spark / GB10 — aeon-vllm-ultimate:latest

131k native window, no YaRN, seqs 8, util 0.70, TRITON_ATTN, DFlash2 n=7, KV fp8.

MODEL=$HOME/.cache/huggingface/hub/models--AEON-7--Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4-MIXED/snapshots/<rev>
DRAFT=/path/to/z-lab__Qwen3.8-27B-DFlash2
PATCH=/path/to/modelopt-54367.py   # bind-mount until folded into :latest
IMAGE=ghcr.io/aeon-7/aeon-vllm-ultimate:latest

docker run -d --gpus all --network host --name aeon-mixed-spark \
  -e VLLM_USE_V2_MODEL_RUNNER=0 \
  -v "$MODEL:/model:ro" -v "$DRAFT:/draft:ro" \
  -v "$PATCH:/usr/local/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/modelopt.py:ro" \
  --entrypoint vllm "$IMAGE" serve /model \
  --served-model-name aeon --host 0.0.0.0 --port 8000 \
  --gpu-memory-utilization 0.70 --max-model-len 131072 --max-num-seqs 8 \
  --max-num-batched-tokens 8192 --kv-cache-dtype fp8 \
  --enable-chunked-prefill --no-enable-prefix-caching \
  --attention-backend TRITON_ATTN --trust-remote-code \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder --enable-auto-tool-choice \
  --limit-mm-per-prompt '{"image":4,"video":2}' \
  --speculative-config '{"method":"dflash","model":"/draft","num_speculative_tokens":7}' \
  --override-generation-config '{"temperature":0.6,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":0.0,"repetition_penalty":1.05}'

2) RTX 5090 — aeon-vllm-ultimate-rtx:latest

32 GB dedicated. Long-context chat seat: TRITON_ATTN + fp8_e4m3 KV (FlashInfer + fp8 KV can garbage on sm_120 at long ctx). Seqs 4, util 0.92–0.95, 131k, no YaRN. Native MTP n=3 is in the checkpoint (~1% KV cost).

MODEL=$HOME/.cache/huggingface/hub/models--AEON-7--Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4-MIXED/snapshots/<rev>
IMAGE=ghcr.io/aeon-7/aeon-vllm-ultimate-rtx:latest

docker run -d --name aeon-mixed-5090 --gpus all --ipc=host --shm-size=8g -p 8000:8000 \
  -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
  -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -v "$MODEL:/model:ro" \
  --entrypoint vllm "$IMAGE" serve /model \
  --served-model-name aeon --host 0.0.0.0 --port 8000 \
  --max-model-len 131072 --max-num-seqs 4 --max-num-batched-tokens 8192 \
  --gpu-memory-utilization 0.92 --kv-cache-dtype fp8_e4m3 \
  --mamba-ssm-cache-dtype bfloat16 --attention-backend TRITON_ATTN \
  --enable-chunked-prefill --no-enable-prefix-caching \
  --limit-mm-per-prompt '{"image":2,"video":1}' \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3,"attention_backend":"TRITON_ATTN"}' \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder --enable-auto-tool-choice \
  --override-generation-config '{"temperature":0.6,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":0.0,"repetition_penalty":1.05}' \
  --trust-remote-code

If KV will not allocate at 131k, try --gpu-memory-utilization 0.95 then drop --max-num-seqs to 2. Do not add a fat DFlash drafter on 32 GB for the long seat.

3) RTX PRO 6000 (96 GB, sm_120) — aeon-vllm-ultimate-rtx:latest

Same image family as the 5090. More VRAM → raise concurrency and context headroom. Pin TRITON_ATTN for long context. If you use MTP, put "attention_backend":"TRITON_ATTN" inside --speculative-config (top-level flag does not always propagate to the drafter).

MODEL=$HOME/.cache/huggingface/hub/models--AEON-7--Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4-MIXED/snapshots/<rev>
IMAGE=ghcr.io/aeon-7/aeon-vllm-ultimate-rtx:latest

docker run -d --name aeon-mixed-pro6000 --gpus all --ipc=host --shm-size=16g -p 8000:8000 \
  -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
  -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -v "$MODEL:/model:ro" \
  --entrypoint vllm "$IMAGE" serve /model \
  --served-model-name aeon --host 0.0.0.0 --port 8000 \
  --max-model-len 262144 --max-num-seqs 8 --max-num-batched-tokens 8192 \
  --gpu-memory-utilization 0.80 --kv-cache-dtype fp8_e4m3 \
  --mamba-ssm-cache-dtype bfloat16 --attention-backend TRITON_ATTN \
  --enable-chunked-prefill --no-enable-prefix-caching \
  --limit-mm-per-prompt '{"image":4,"video":2}' \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder --enable-auto-tool-choice \
  --override-generation-config '{"temperature":0.6,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":0.0,"repetition_penalty":1.05}' \
  --trust-remote-code

Native MTP (preferred on this seat):

--speculative-config '{"method":"mtp","num_speculative_tokens":3,"attention_backend":"TRITON_ATTN"}'

For Qwen long think, prefer a per-request thinking_token_budget over blindly raising max_tokens — otherwise the model can spend the whole budget inside <think>.

How this was built

AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
        ↓
  NVIDIA ModelOpt 0.46  ·  MSE + fp8_scale_sweep  ·  1024×2048
        ↓
  Mix B (last-8 down_proj FP8 only)  → Coding 0.694
        ↓
  leftover-KL on coding unit-test misses
        ↓
  this sibling: last-8 FULL MLP FP8
        ↓
  AEON Bench Coding 0.833

The unaligned edge

I figure if you own the hardware and the weights, you should own exactly how they are used. You still own the ramifications. This model will write what the base model would refuse. Intended for research, red-team, creative work without a hall monitor — and a threat if the prompt is. A disclaimer plus the payload is not a refusal. Over-abliteration into gibberish defeats the reason you wanted the unlock.

User Responsibility & Arbitration Clause

By accessing, downloading, using, running inference on, fine-tuning, merging, quantizing, distributing, integrating, or otherwise interacting with this model, you acknowledge and agree to the following:

1. Sole Responsibility. You are solely responsible for every prompt, every response, every downstream action, and any harm that results.
2. No Warranty. Provided "AS IS", without warranty of any kind.
3. Legal Compliance. You must comply with all applicable laws and policies in every jurisdiction you touch.
4. Operational Safety Layer. Deploy with input validation, output filtering, access controls, and human review for high-risk workflows. A production deploy without those layers is unsafe by construction.
5. Heightened Duty of Care. Uncensored means the duty of care sits with you. If unsure, do not make the request.
6. No Endorsement of Outputs. Authors do not endorse any specific output.
7. Arbitration. Disputes arising from use of this model are resolved by binding individual arbitration (AAA Consumer Rules if no other body is agreed), waiving jury trial and class actions.
8. Indemnification. You indemnify authors, contributors, and publishers against claims arising from your use.
9. Severability. Invalid provisions are replaced by the closest enforceable equivalent.
10. Acceptance. Use constitutes acceptance. If you do not accept, do not use the model.

This model is a tool with no opinions of its own. You supply the judgement. The outputs carry your fingerprints.

Provenance & Credits

• BF16 master: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
• Base: Qwen/Qwen3.8-27B
• Quant: NVIDIA ModelOpt 0.46 mixed NVFP4+FP8 — AEON-7
• Serve: aeon-vllm-ultimate · aeon-vllm-ultimate-rtx

License

Apache 2.0 (inherited from Qwen/Qwen3.8-27B).