aeon-7/qwen3.6-27b-aeon-ultimate-uncensored-ddtree

This is an experimental research repository aimed at making a large AI model run faster on specialized hardware. The model in question is Qwen3.6-27B AEON Ultimate, a hybrid AI model that combines standard attention processing with recurrent, state-keeping layers, making it more complex than typical language models. The hardware target is DGX Spark and GB10 systems, which are high-end AI computing platforms. The core research problem here is called speculative decoding, a technique for speeding up AI text generation. Normally, a model generates one token, meaning one word piece, at a time, with a cheap draft model proposing candidates and an expensive target model verifying them. Standard speculative decoding follows a single chain of guesses. DDTree, short for Decision-tree Decoding, instead explores a tree of alternative branches at once, so if the top guess turns out wrong, another branch may still be accepted without wasting the verifier's effort. Applying DDTree to Qwen3.6 is technically difficult because the model's recurrent layers mean each branch of the tree needs its own copy of the running internal state, something standard speculative decoding setups do not need to handle. The README is candid that this is unfinished work, listing what has been tried, what appears to work, and what still breaks, alongside raw benchmark results and a working experimental container image. The project ships as a Docker container built on top of the vLLM serving framework, published with several run modes: a safe research mode with conservative limits, a mode that behaves like the stable production version, and an unsafe full research mode meant only for people developing the underlying kernels. Running any of it requires a GPU-equipped DGX Spark or GB10 machine and the associated model weights downloaded separately. The repo is intended for AI researchers and infrastructure engineers, not production use, and is released under the Apache 2.0 license.