I build deployment-ready open releases for next-gen hardware — NVFP4-quantized abliterated LLMs (Gemma 4, Qwen 3.6, Nemotron 3) on NVIDIA DGX Spark (GB10 / Blackwell / sm_121a), DFlash + EAGLE speculative decoding, Apple Silicon MLX builds for M-series Macs, a real-time voice-AI stack, and a complete agent-driven AI media production toolchain.

Everything below is public, MIT/Apache-licensed, and reproducible — Docker stacks, pre-built vLLM images, deployment guides, and benchmark numbers included. Model weights live on 🤗 Hugging Face.

Open-source AI-driven media production toolchain. Five focused repositories — each generating one kind of media (music, radio drama, music video, cinematic video, or the base ComfyUI stack), all designed for AI agents through skill MD files and CLI scripts. No node-graph wrangling.

Repo	What it does	★
comfyui-aeon-spark	Bleeding-edge ComfyUI for DGX Spark (CUDA 13 + SageAttention v3 + NVFP4 + 14 custom-node packs + Flux 2 Dev / LTX 2.3 22B / ACE-Step v1.5 XL Turbo pre-bundled). Foundation for every other repo in this section.
aeon-music-maker	ACE Step 1.5 XL music generation with dynamics-preserving mastering chain (HPF → EQ → tape sat → LUFS gain-match → true-peak ceiling). FLAC-lossless output, auto-detected mastering presets, CLI-driven.
aeon-radio-drama	Full-pipeline radio drama / audiobook production — dialogue (Qwen3-TTS) + music (ACE Step) + SFX (MMAudio / Stable Audio Open / ACE) + sidechain mix in one command. Three-Lock voice persistence. Bundles standalone music_maker.py + sfx_maker.py for one-shot music or SFX generation.
aeon-music-video	Audio-reactive music video builder. librosa-driven beat / onset / RMS / spectral-centroid detection drives ffmpeg filter chains for synced visual effects. CPU-only — no GPU, no ComfyUI, no model downloads required.
aeon-movie-maker	Fast cinematic video via LTX 2.3 22B. Single clips, full screenplays with character continuity (last-frame carry-forward + per-character seed offsets), and sidechain-mixed final cuts. CLI-tunable LoRA strengths + saturation troubleshooting guide.

git clone https://github.com/AEON-7/<repo>     →   ./setup.sh        →   start using
                                                       │
                                                       ▼
                                              copy .env.example → .env
                                              edit COMFYUI_URL etc.
                                                       │
                                                       ▼
                                          python scripts/<tool>.py ...
                                                       │
                            (later, when upstream updates) ▼
                                                ./sync.sh
                                              (preview → confirm → pull → refresh)

Every tool that uses ComfyUI supports two execution modes, documented per-repo:

• Mode A — Local: CLI runs on the same machine as ComfyUI. Just python scripts/<tool>.py ....
• Mode B — Remote: ComfyUI on a GPU box (DGX Spark, headless server). Either invoke the CLI over SSH (ssh user@gpu-host 'cd repo && python ...') or hit the remote ComfyUI HTTP API directly via SSH tunnel or --listen 0.0.0.0.

aeon-movie-maker has additional constraints documented (I2V + screenplay carry-forward needs filesystem-level access — pure HTTP-only remote works for T2V single clips only).

Real-time speech AI for DGX Spark. Three composable sidecars turn any Spark into a voice agent host: pair two OpenAI-compatible audio endpoints (TTS + ASR) with the Matrix WebRTC bridge to dial your AI directly from any Matrix client — then compose all three into fully-embodied AI personas with create-agentic-personas. End-to-end voice turn: ~2.1 s on Spark.

Repo	What it does	★
qwen3-tts-server	OpenAI-compatible /v1/audio/speech server backed by Qwen3-TTS-12Hz-1.7B-VoiceDesign. CUDA + bf16 + flash-attn 2 (sm_120 wheel). RTF 1.30× hot path (1.48 s synthesis for ~2 s of speech). Pre-built ghcr image, deploy scripts covering 5 model variants (VoiceDesign / CustomVoice / Base @ 1.7B & 0.6B).
qwen3-asr-server	OpenAI-compatible /v1/audio/transcriptions server — Qwen3-ASR-0.6B served by vLLM. 30 spoken languages + 22 zh dialects. RTF 16× hot path (120 ms transcription for 2 s of audio). Pre-built ghcr image, deploy scripts for 0.6B / 1.7B variants.
matrix-voip-agent	Headless Matrix WebRTC voice agent — auto-answers VoIP calls and bridges audio to any AI agent via PipeWire. The recommended bridge for AI-on-Matrix-VoIP: combine with any Matrix homeserver (Synapse / Conduit) and the two sidecars above to dial your AI directly from any Matrix client.
create-agentic-personas	Build fully-embodied AI personas on OpenClaw + Matrix — each with a chat identity, a knowledge corpus (RAG), a cloned-or-designed Qwen3-TTS voice, and a live WebRTC call line. Composes the three sidecars above into a one-command-per-persona roster builder: secret-free templates, a new-persona.sh scaffold, and a create-agentic-persona agent skill so an agent can spin up new personas itself.

The three voice sidecars + the Qwen3.6-27B AEON Ultimate MTP-XS vLLM main on one Docker bridge = a complete sub-3-second voice agent. Latency budget (measured, hot path on DGX Spark):

Each repo ships with a README.md, agents.md (autonomous bring-up runbook), docs/MODELS.md (variant catalog), docs/ARCHITECTURE.md (full topology), and docs/INTEGRATIONS.md (Matrix + OpenAI SDK + OpenWebUI + Home Assistant + raw HTTP).

Abliterated Gemma 4 deployments at NVFP4 precision (4-bit weights) for NVIDIA DGX Spark / Blackwell GPUs — quantized weights, EAGLE speculative-decoding drafters, and a validated DFlash serving container that more than triples single-stream throughput.

Repo	Model	Architecture	Description	★
Gemma-4-31B-Uncensored-NVFP4-DFlash	Gemma 4 31B Deckard Heretic	Serving container + DFlash	Validated vLLM container pairing the 31B DECKARD NVFP4 weights with the official z-lab DFlash drafter. 3.5× single-stream (11 → 39 tok/s) and up to 427 tok/s aggregate @ c=32 on DGX Spark — with reasoning, tool calling, vision/video input, and structured output fully intact.
Gemma-4-31B-DECKARD-HERETIC-Uncensored-NVFP4	Gemma 4 31B DECKARD HERETIC	Dense, thinking	NVFP4-quantized abliterated 31B dense reasoning model. AWQ_FULL + SVDQuant variants. 🤗 weights
Gemma-4-E4B-DECKARD-HERETIC-Uncensored-NVFP4	EAGLE drafter for 31B DECKARD	Speculative decoding	EAGLE E4B speculative-decoding drafter for the 31B DECKARD HERETIC. 🤗 weights
Gemma-4-26B-A4B-it-Uncensored-NVFP4	Gemma 4 26B A4B-it	MoE	NVFP4-quantized 26B MoE. 50 tok/s single, 1430 tok/s aggregate on DGX Spark. 🤗 weights
Gemma-4-E4B-it-Uncensored-NVFP4	EAGLE drafter for 26B MoE	Speculative decoding	EAGLE E4B speculative-decoding drafter for the Gemma 4 26B MoE. NVFP4 AWQ. 🤗 weights
supergemma4-26b-abliterated-multimodal-nvfp4	SuperGemma4 26B Multimodal	Multimodal · 🗄️ archived	NVFP4 AWQ full quantization of SuperGemma4-26B-Abliterated-Multimodal — pre-built vLLM container + patches included. Archived; kept public for reference and reproducibility. 🤗 weights

The AEON catalog comes to the Mac. Metal-accelerated MLX builds of abliterated Gemma 4 — fully multimodal (text + image + audio), OpenAI-compatible, running host-native on any M-series machine. No CUDA required, no Docker GPU passthrough games.

Repo	What it does	★
gemma4-aeon-abliterated-mlx-toolkit	Apple-Silicon toolkit + OpenAI-compatible server for the Gemma-4-12B AEON Abliterated MLX quant grid: near-lossless MLX-8bit (13.4 GB) flagship and compact MLXFP4 (9.3 GB) for 16 GB Macs. One-paste uv quickstart boots a multimodal mlx_vlm.server on a fresh Mac — verified image description and speech transcription through the API. Optional MTP speculative decoding (~1.1–1.2× faster, output-identical). Benchmarked on M4 Pro 48 GB.

Grab the weights: 🤗 MLX-8bit · 🤗 MLXFP4 · 🤗 K4-BF16 source

The flagship line. Lossless abliteration of Qwen 3.6 with hardware NVFP4 quantization — dense 27B and 35B MoE — combined with DFlash speculative decoding for serious single-stream throughput on DGX Spark, plus an open research track pushing speculative decoding for hybrid-attention models forward.

Repo	Model	Architecture	Description	★
Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash	Qwen 3.6 27B AEON Ultimate Uncensored	Dense	The most-starred release in the catalog. Lossless abliteration with NVFP4 hardware quantization — BF16 (51 GB) + NVFP4 (26 GB) deployment guide, docker-compose, and QuickStart. The production serving path for Qwen 3.6 on Spark. 🤗 weights
Qwen3.6-NVFP4-DFlash	Qwen 3.6 35B-A3B-heretic	MoE	NVFP4 + DFlash speculative decoding on DGX Spark (GB10 / sm_121a). Source-built vLLM image + 7 patches + comprehensive deployment guide. 🤗 weights
Qwen3.6-27B-AEON-Ultimate-Uncensored-DDTree	Qwen 3.6 27B AEON Ultimate Uncensored	🔬 Experimental research track	DDTree-on-vLLM for hybrid-attention Qwen 3.6 — tree verification, branch-state replay, Gated DeltaNet state handling, fused branch attention. Intentionally candid lab notes: what's been tried, what works, what still breaks, and where the next breakthrough likely lives. Use the DFlash repo above for production.

NVIDIA Nemotron deployments for Blackwell-class hardware.

Repo	Model	Architecture	Description	★
Nemotron-3-Nano-Omni-AEON-Ultimate-Uncensored	Nemotron 3 Nano Omni	12-D abliterated multimodal	BF16 + NVFP4 multimodal reasoning model for DGX Spark / Blackwell. Source-built vLLM v0.20.0 image + 4 patches + benchmark + deployment guide. 🤗 weights

The engine room: the unified serving image that runs the whole catalog, plus the speculative-decoding, KV-cache-compression, and quantization building blocks underneath it.

Repo	What it does	★
vllm-ultimate-dgx-spark	⭐ AEON vLLM Ultimate — the current flagship serving image (ghcr.io/aeon-7/aeon-vllm-ultimate). vLLM 0.22.1 + Triton NVFP4 KV cache (PR #44389 cherry-pick — up to 3× KV capacity) + TurboQuant K8V4 4-bit KV compression + native DFlash / EAGLE3 via --speculative-config + 4 idempotent sm_121a runtime patches. One image serves the entire AEON model catalog on DGX Spark and RTX 50-series Blackwell.
vllm-dflash	The original DFlash vLLM image for DGX Spark — Plug & Play Block-Diffusion Speculative Decoding with NVFP4, sm_121a kernels, and Qwen-targeted optimizations. Start with AEON vLLM Ultimate (above) for new deployments.
turboquant	Near-optimal KV-cache quantization for LLM inference (3-bit keys, 2-bit values) with Triton kernels + vLLM integration. This fork carries the CUDA-graph-safe QJL _POWERS fix that lets TurboQuant boot under CUDA graph capture — bundled into AEON vLLM Ultimate as --kv-cache-dtype tq_k8v4.
Model-Optimizer	Tracking fork of NVIDIA's unified model-optimization library — quantization, pruning, distillation, speculative decoding — for TensorRT-LLM / TensorRT / vLLM deployment. The quantization workhorse behind every NVFP4 release on this page.
modelopt-fast-moe	MoE-targeted quantization + AWQ calibration tooling. NVFP4 routing, expert-aware modelopt.

Every public container on ghcr.io/aeon-7 — built, validated, and mapped to the repo that documents it. All images: docker pull ghcr.io/aeon-7/<image>.

Side projects, tools, and infrastructure that aren't model deployments but might be useful.

Repo	What it does	★
unifi-ai-network-management	Agent-ready UniFi / Ubiquiti network management skill and tooling: safe API key setup, backup/restore helpers, OpenClaw + Hermes install paths, and operational playbooks for diagnostics, security events, clients, APs, switches, VLANs, and Wi-Fi automation.
cosmic-mind	Security-and-resiliency-focused deployment of the Quartz web app. A place to build your second mind and share it.
regex-builder	Simple and elegant RegEx builder.
quartz	Fast batteries-included static-site generator that transforms Markdown into fully functional websites. Fork — the upstream base for cosmic-mind.

If any of these releases have been useful to you, tips are deeply appreciated — they go directly toward more compute, more models, and more open releases. Scan a QR with your wallet, or click any address below to copy.

₿ Bitcoin (BTC) bc1q09xmzn00q4z3c5raene0f3pzn9d9pvawfm0py4	Ξ Ethereum (ETH) 0x1512667F6D61454ad531d2E45C0a5d1fd82D0500
◎ Solana (SOL) DgQsjHdAnT5PNLQTNpJdpLS3tYGpVcsHQCkpoiAKsw8t	ⓜ Monero (XMR) 836XrSKw4R76vNi3QPJ5Fa9ugcyvE2cWmKSPv3AhpTNNKvqP8v5ba9JRL4Vh7UnFNjDz3E2GXZDVVenu3rkZaNdUFhjAvgd

Ethereum L2s (Base, Arbitrum, Optimism, Polygon, etc.) and EVM-compatible tokens can be sent to the same Ethereum address.

• 🌐 Open an issue on any repo for questions, bug reports, or feature requests
• 🤗 Model weights, quant grids, and drafters live on Hugging Face → AEON-7
• 📜 Most releases include a deployment guide + benchmark numbers — start there