Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Fully uncensored, capability-enhanced abliteration of Qwen3.6-27B. NVFP4 + z-lab DFlash speculative decoding (n=12) on the unified ghcr.io/aeon-7/aeon-vllm-ultimate:latest container, tuned for long-context draft acceptance on DGX Spark. 6 HF variants (BF16/NVFP4/MTP/MTP-XS), docker-compose, and QuickStart.
| Date | Stars |
|---|---|
| 2026-07-24 | 424 |
| 2026-07-25 | 424 |
| 2026-07-28 | 424 |
| 2026-07-30 | 424 |
| 2026-08-06 | 424 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<div align="center">
# Qwen3.6-27B-AEON-Ultimate-Uncensored
### Lossless abliteration · Capability-enhanced · NVFP4 hardware-quantized for Blackwell
[-yellow?logo=huggingface)](https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16)
[-yellow?logo=huggingface)](https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-NVFP4)
[](https://github.com/AEON-7/vllm-ultimate-dgx-spark/pkgs/container/aeon-vllm-ultimate)
[](LICENSE)
[](https://github.com/AEON-7/AEON-7#-support-the-work)
**Refusals: 0 / 100** · **KL vs base: 0.000492** · **Compression: 49 %** · **Capability: enhanced**
</div>
---
## TL;DR
A **fully uncensored, capability-enhanced** abliteration of [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B), produced over **72 hours of continuous research** drawing on hundreds of parallel AI research agents, the industry's best published methodologies, custom in-house techniques, and yet-unreleased pre-public branches of next-generation abliteration software.
## Quickstart (DGX Spark / GB10) — copy-paste
> ⚠️ **This container is DGX Spark / GB10 only — it is not portable.** `aeon-vllm-ultimate` is built for the Spark's **ARM64 CPU** and GB10's **`sm_121a` NVFP4 kernels**. On an **x86 / amd64 host it will not start** (`/usr/bin/bash: cannot execute binary file` — wrong CPU arch); on a **non-Blackwell GPU** (RTX 4090 / L40S = Ada `sm_89`, A100/H100 = Ampere/Hopper) the **NVFP4 path can't run** (NVFP4 needs Blackwell). **For any hardware other than the DGX Spark, do not pull this container** — instead run the portable [HuggingFace weights](#model-variants) on your *own* vLLM build, picking the right variant + spec-decode method for your GPU from the [Hardware compatibility matrix](#hardware-compatibility-matrix). *(An x86 + Blackwell image may come in a future release.)*
One block: pull the container, pull this model (fresh, the `-Multimodal-NVFP4-MTP-XS` body — see [Model Variants](#model-variants)), pull the DFlash drafter (fresh), then serve with the vetted DGX Spark flags.
```bash
# 1) Pull the unified AEON vLLM container (vLLM 0.23.0, sm_121a, DFlash)
docker pull ghcr.io/aeon-7/aeon-vllm-ultimate:latest
# 2) Pull this model (fresh)
huggingface-cli download AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP --local-dir ./aeon-model
# 3) Pull the DFlash drafter (fresh — z-lab pushes updates; always re-pull)
huggingface-cli download z-lab/Qwen3.6-27B-DFlash --local-dir ./aeon-drafter
# 4) Serve (ENTRYPOINT is /bin/bash → pass --entrypoint vllm, then serve ...)
docker run --gpus all --ipc host --network host \
-v ./aeon-model:/model:ro -v ./aeon-drafter:/drafter:ro \
--entrypoint vllm ghcr.io/aeon-7/aeon-vllm-ultimate:latest \
serve /model \
--quantization modelopt \
--mamba-cache-dtype float32 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder \
--enable-auto-tool-choice \
--limit-mm-per-prompt '{"image":4,"video":2}' \
--mm-encoder-tp-mode data \
--max-num-seqs 16 \
--max-num-batched-tokens 16384 \
--gpu-memory-utilization 0.70 \
--enable-chunked-prefill \
--enable-prefix-caching \
--trust-remote-code \
--speculative-config '{"method":"dflash","model":"/drafter","num_speculative_tokens":10}'
```
> **`--gpu-memory-utilization 0.70` is the safe default.** The Spark's 128 GB is *unified* (CPU + GPU share one LPDDR5X pool), so over-allocating page-thrashes — which shows up as slowdowns, agent loops, and even engine crashes, not an OOM error. 0.70 leaves headroom for co-Excerpt of 67,227 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:206e91aa49d2df74, topic:llm, topic:qwen
matched fp:206e91aa49d2df74, topic:vllm, desc:speculative decoding
matched fp:206e91aa49d2df74, topic:quantization, readme:quantization, readme:quantized