AI Software

IWhy

The most capable models in the world were built to be rented. Rented intelligence has a landlord. The terms and the price can change, the answers can be shaped by rules you never agreed to, and your conversations live on someone else's disk.

We think a person should be able to own this the way they own a workshop. A model on your own machine answers to you, and no meter is running.

Owning it is harder than renting it. The models are large, the hardware is new, and the software underneath changes every month. The work of this pillar is to close that gap and publish every step.

IIWhat we build

Open releases that run on hardware a person can buy and put on a desk.

• Abliterated language models at NVFP4 precision for NVIDIA DGX Spark and other Blackwell-class GPUs: Gemma 4, Qwen 3.6, and Nemotron 3, with weights on Hugging Face and guides on GitHub.
• Apple Silicon builds of the same catalog in MLX, with an OpenAI-compatible multimodal server for M-series Macs.
• AEON vLLM Ultimate, one serving image for the whole catalog, with NVFP4 KV cache, TurboQuant KV compression, and DFlash and EAGLE speculative decoding.
• A real-time voice and video stack: streaming speech servers, a Matrix VoIP bridge, and a persona builder that turns them into an agent you can call.
• An agent-driven media production toolchain on ComfyUI: music, radio drama, music video, and cinematic video from the command line.
• AEON Bench, an open, attested benchmark with a public leaderboard at aeon-bench.com.

Everything is MIT or Apache licensed, with Docker stacks, pre-built images, and benchmark numbers.

IIIPrinciples

Reproducible or retracted. Every release ships with the recipe that produced its numbers. If you cannot reproduce them, the release is wrong, not you.

Run where you are. The target is the machine on your desk, not a cluster you have to lease.

Say what the model is. Our releases are abliterated: the refusal behavior trained into the base model has been removed, so the model answers the question it was asked. We say so on every model card. Whoever runs the model is responsible for how it is used, as with any tool.

Measure in the open. AEON Bench attests every ranked run by re-fetching the weights and re-verifying every hash before a score counts.

Credit the source. Every release names the base model and the work it builds on.

IVCurrent work

The Qwen 3.6 27B AEON Ultimate release is the production serving path on DGX Spark, with BF16 and NVFP4 weights and a DFlash serving container. The Gemma 4 31B DFlash container more than triples single-stream speed, from 11 to 39 tokens per second, with reasoning, tool calling, and vision intact. The Gemma 4 26B mixture-of-experts release reaches 50 tokens per second single-stream.

On Apple Silicon, Gemma 4 12B and Qwen 3.6 27B run as MLX builds with text, image, and audio input through one server.

The voice stack starts speaking about one second into a turn on a single Spark, measured end to end.

AEON Bench is live. A DDTree research track on speculative decoding is published as candid lab notes, failures included.

VHow to take part

Pull a model and run it. Open an issue with what breaks.

Benchmark your own hardware with Aeon Bench Pod and submit an attested run to the leaderboard.

Contribute patches, guides for new hardware, or a request for a model you want quantized.

Sign in to the Forge to follow the work as it happens.