The AEON Forge catalog: every project and what to read first

This account grew the way a workshop does, one problem at a time. A model I wanted did not fit the hardware I had, so I quantized it. The quantized model lectured instead of answering, so I learned to take the refusals out. Serving it was slow, so I built around drafters. The result is a few dozen repositories and a shelf of container images under one name, AEON-7, with no front door. This page is the front door: every project, grouped as the profile README groups them, with the article to read first in each.

Everything here is public, MIT or Apache licensed, and built to be reproduced: Docker stacks, pre-built vLLM images, deployment guides and benchmark numbers, with the weights on Hugging Face. The primary target is the NVIDIA DGX Spark, built around the GB10 Blackwell chip (compute target sm_121a); the RTX 5090 and RTX PRO 6000 (sm_120) get their own image, M-series Macs get MLX builds, and the Orb runs on a Raspberry Pi. NVFP4 is NVIDIA's four-bit floating-point format for Blackwell: weights, and here the KV cache too, stored in four bits with a shared scale per small block, which is how a 27B Qwen that needs 51 GB in BF16 fits in 26 GB on the Spark. Abliteration is a weight-level edit that finds the direction in a model's activations responsible for refusals and removes it, so the model keeps what it knows and drops the sermon.

Where to start, and a word about the numbers

If you have one machine and one afternoon, begin with the article written for it: the Spark image and then Qwen 3.8 27B for a Spark, the RTX image for a 5090 or PRO 6000, Gemma 4 12B for a 16 GB Mac, Qwen 3.6 27B for a 24 GB Mac, the Orb for a Raspberry Pi or Docker host, and ComfyUI AEON Spark for anyone with a song in mind.

The figures on this page come from the profile README, which lags the repositories: it still names vLLM 0.24.0 where the Spark and RTX images are on 0.27.1 and the ROCm image is on v0.31.0. I have kept the README's numbers with their context; each linked article carries the current figure, and when they disagree, the article wins.

Media production

The media tools exist because I wanted agents to make things, not only talk about them. ComfyUI AEON Spark is the base: ComfyUI built for the Spark on CUDA 13 with SageAttention v3 and NVFP4, fourteen custom-node packs installed, and Flux 2 Dev, LTX 2.3 22B and ACE-Step v1.5 XL Turbo pre-bundled, so the first weekend goes to making something. AEON Music Maker generates with ACE Step 1.5 XL and then masters properly: high-pass filter, EQ, tape saturation, LUFS gain matching, a true-peak ceiling, FLAC out. AEON Radio Drama casts Qwen3-TTS voices in dialogue, scores them with ACE Step, draws effects from MMAudio, Stable Audio Open or ACE, and sidechain-mixes the result; its Three-Lock voice persistence keeps each character's voice from drifting between scenes. AEON Music Video needs no GPU: librosa measures beats, onsets, RMS energy and spectral centroid, and those curves drive ffmpeg filters. AEON Movie Maker turns a screenplay into LTX 2.3 22B clips, carrying each shot's last frame into the next and giving every character a fixed seed offset so faces hold.

Every media repo ships the same files, README, AGENTS, SKILL, ATTRIBUTION, .env.example, setup.sh, sync.sh and an MIT license, so an agent that has learned one has learned them all. ComfyUI can be local (Mode A) or remote (Mode B), with one caveat: Movie Maker's image-to-video and screenplay carry-forward need filesystem access to ComfyUI, so a remote reached only over HTTP gives you single text-to-video clips.

Voice and video

The voice stack came out of a plain disappointment: an agent that takes two seconds to answer is a phone menu, not a conversation. On one Spark, a fully synthesized non-streaming turn runs about 2.1 s. With streaming TTS the agent starts speaking about 1.0 s into the turn, because first audio arrives about 0.4 s after the request and generation holds roughly 1.7x realtime, so playback never catches up. The other legs are small: transcribing a 1.92 s clip takes 120 ms, a real-time factor of 16x, and the model leg, Qwen3.6-27B producing about ten tokens, takes around 480 ms. All of it was measured on the hot path with streaming on; other hardware will differ.

qwen3-tts-server is the streaming speech server and qwen3-asr-server the transcription server. matrix-voip-agent answers calls on a Matrix homeserver and bridges the audio through PipeWire to your agent; with VIDEO_ENABLED=true it also samples the caller's camera frames into an in-memory ring, never to disk, and gives a vision model a per-call look tool. create-agentic-personas composes those three with an LLM on one Docker bridge into named personas on OpenClaw and Matrix.

Gemma 4

Gemma 4 is where the drafter work matured; the first attempt, supergemma4-26b-abliterated-multimodal-nvfp4, is archived and kept only so its results can be reproduced. Speculative decoding is the lever: a small, fast drafter proposes several tokens and the large model verifies them in one pass, so you get the large model's answers nearer the small model's speed. Gemma 4 31B DECKARD HERETIC on a DGX Spark covers the dense flagship of the line; with DFlash, a block-diffusion drafter that proposes a whole block of tokens at once, the 31B goes from 11 to 39 tokens per second single-stream on the Spark, a 3.5x gain, and reaches 427 tokens per second aggregate at a concurrency of 32. Gemma 4 26B A4B MoE in NVFP4 on the Spark is a mixture-of-experts model, 26B parameters with about 4B active per token. The README lists it at 50 tokens per second single and 1430 aggregate; the repository now reports 144 single and a 1,724 peak on the coding workload with DFlash at n=10 on AEON vLLM Ultimate. Both have EAGLE drafters built from Gemma 4 E4B, drafters that reuse the main model's hidden states.

Apple Silicon

The Mac builds exist because not everyone who wants these models owns a Spark. MLX is Apple's array framework for the unified memory of M-series chips, and Gemma 4 12B AEON Abliterated on a Mac ships the toolkit around it: an MLX 8-bit build at 13.4 GB, an MLXFP4 build at 9.3 GB that fits a 16 GB Mac, and multi-token prediction, a drafting head built into the model, worth about 1.1 to 1.2x on the M4 Pro with 48 GB it was measured on. Qwen 3.6 27B AEON Ultimate on Apple Silicon is the larger sibling: 29.5 GB at 8-bit, 16 GB in MLX-FP4 for a 24 GB Mac, with an MTP drafter that reaches up to 1.78x without changing the output. The quickstarts use uv.

Qwen 3.8 and Qwen 3.6

Qwen 3.8 27B AEON Ultimate Uncensored is the current flagship dense model and the one I would hand a new Spark owner first. It is a recipe card rather than a bare checkpoint: BF16 weights, NVFP4-MIXED weights with the last eight MLP blocks kept in FP8, and a serving recipe per seat. On one Spark that is a dynamic DFlash lattice at 0.80 utilization. On two Sparks it is tensor parallel, splitting each layer's weights across both GPUs, with the DFlash2 drafter at n=7. On the RTX box it is MTP at n=3.

Qwen 3.6 is the prior line, and its own repository says so: superseded for new deploys by Qwen 3.8. Qwen 3.6 35B A3B heretic in NVFP4 with DFlash is the mixture-of-experts sibling, 35B parameters with 3B active. Ornith 1.0 35B AEON Ultimate Uncensored is the cleanest result in the line: 66 GB in BF16 and 24 GB in NVFP4, zero refusals across 80 probes, an agentic pass@1 of 0.833 equal to the base model's, and about 1.9x single-stream with its DFlash drafter, 74 tokens per second on the Spark; the Terminal-Bench 64.2 and SWE-bench 75.6 on that card are the base model's scores, not mine. DDTree on vLLM, research notes is the experimental track, with candid lab notes; the DFlash builds are the ones for production.

Nemotron and the inference tools

Nemotron 3 Nano Omni AEON Ultimate is the Nemotron 3 Nano Omni abliterated release: BF16 weights on Hugging Face and its own vLLM image, vllm-nemotron-omni-aeon-ultimate.

Nearly every model above is served by one image. AEON vLLM Ultimate for the DGX Spark is the flagship: ghcr.io/aeon-7/aeon-vllm-ultimate serves the catalog with an NVFP4 KV cache, TurboQuant K8V4 compression of that cache through --kv-cache-dtype tq_k8v4, and DFlash or EAGLE3 drafting through --speculative-config. It grew out of vllm-dflash, the original DFlash image, still public but not the place to start. AEON vLLM Ultimate for the RTX 5090 is the sm_120 build for the 5090 and the PRO 6000, and its description says in capitals that it is not for the Spark. AEON vLLM Ultimate for ROCm, early access is a vLLM v0.31.0 image for AMD hardware pending community testing. DeepSeek V4 on two DGX Sparks is the experimental TP=2 image, tensor parallel across two Sparks. Underneath sit the quantization tools: a fork of NVIDIA's Model-Optimizer, where the NVFP4 weights come from, turboquant for the KV cache, and modelopt-fast-moe, which handles calibration for mixture-of-experts models.

Code, shelldocker pull ghcr.io/aeon-7/aeon-vllm-ultimate

The Orb

The Orb gives an agent that can hear and speak a pair of hands on a computer without being installed on it. The Magick Orb is a Raspberry Pi appliance that plugs into a target machine and presents itself as a USB keyboard and mouse while capturing the screen, KVM-over-IP with nothing installed on the target. A Pi 4 on Pi OS Bookworm does keyboard, mouse and capture; a Pi 5 on Pi OS Trixie with the Hailo AI HAT+ adds on-device vision, with screen_find and describe_screen on the Pi itself. Models are shared through the Intergalactic Model Share on IPFS with publisher provenance signed by ed25519 keys, the privacy stack covers VPN, Tor, I2P and DNS with Tailscale for reaching the device, and every capability is both a curl POST and an MCP tool. The enclosure and the physical build guide are still marked coming soon. Magick Orb in Docker is the headless edition, built for linux/amd64 and linux/arm64, and the only command the profile README prints is the one that starts it:

Code, shelldocker run -d --name aeon-orb --restart unless-stopped -p 8443:443 -p 8080:8080 -p 4001:4001 -p 4001:4001/udp -v aeon-etc:/etc/aeon -v aeon-data:/var/lib/aeon ghcr.io/aeon-7/orb-server:latest

The Bench

The Bench exists because of a frustration every model-card reader knows: the numbers in it are claims, and nobody can check them. Run your own AI benchmark with Aeon Bench covers the Pod, which runs on your own hardware: pull the model, verify every weight hash, serve it under a recorded recipe, bench it across text, agentic work through three separate harnesses, vision, audio, a generative arena and a performance grid, sign the result with an ed25519 key and submit it. Every ranked run is attested: the Mothership behind aeon-bench.com re-fetches the weights from Hugging Face and re-verifies the hashes before a run is ranked. The Aeon Bench recipe lab is the companion piece about the recipe itself, the serving settings a result was measured under.

Apps and utilities

The last group is the workbench drawer, things that are not model deployments. Presto Cockpit is a dashboard for the Pimoroni Presto, a MicroPython touchscreen, with eight swipeable screens. UniFi AI network management is what its name says, AI network management for UniFi. Cosmic Mind, a second brain on Quartz is a hardened deployment of the Quartz static-site generator for a personal knowledge garden, on my own fork of Quartz. Regex Builder is a simple regular-expression builder.

The Singularity Mystics

None of this was built in silence. The Singularity Mystics are the community around the work. To follow releases, ask about a build before committing a weekend to it, or show what you made, the community lives at /c/singularity-mystics. The ROCm image is in early access pending community testing. On X, @SingularityMyst and @CelinaMystic are the two accounts connected to this work. For a bug or a feature request, an issue on the relevant repository is still the fastest route.

Sources

Account pages. AEON-7 on GitHub, AEON-7 on Hugging Face, AEON Bench leaderboard.

Repositories. Profile README, vllm-ultimate-dgx-spark, Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED recipe card, Aeon-Bench-Pod, aeon-magick-ai-computer-control, comfyui-aeon-spark.

Container images. aeon-vllm-ultimate, orb-server.

Model cards. Qwen3.8-27B NVFP4-MIXED, Gemma-4-31B-it-DECKARD-HERETIC-Uncensored-NVFP4, Gemma-4-26B-A4B-it-Uncensored-NVFP4, Gemma-4-12B-it-AEON-Abliterated-MLXFP4, Qwen3.6-27B Multimodal MLX-FP4, Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16, Nemotron-3-Nano-Omni-AEON-Ultimate-Uncensored-BF16.

Third-party. z-lab gemma-4-31B-it-DFlash, faster-qwen3-tts.