papers/zoo-per-llm-chains/sections/06-zen-models.tex

\section{The \Zen{} Model Family}
\label{sec:zen}

\subsection{Canonical deployment}

The \Zen{} family\footnote{Source repositories: \texttt{\textasciitilde/work/zen} (training), \texttt{huggingface.co/zenlm} (releases).} is the canonical \Zoo{} per-LLM chain deployment. Each released \Zen{} model has a sovereign chain identified by the canonical id \texttt{zenlm/<model>@<major>}.

The lineup below is the current \texttt{zen4} generation deployed at chain genesis (2025-12-25) plus the in-development \texttt{zen5} line whose chains are reserved.

\subsection{Small line (Qwen3.5 base, Gated DeltaNet, 262K context)}

\begin{longtable}{p{3.5cm}rrlp{4cm}}
\toprule
\textbf{Model} & \textbf{Total} & \textbf{Active} & \textbf{License} & \textbf{Notes} \\
\midrule
\endhead
\texttt{zen4-nano} & 0.8B & 0.8B & Apache 2.0 & abliterated; edge \\
\texttt{zen4-micro} & 2B & 2B & Apache 2.0 & abliterated; mobile \\
\texttt{zen4-mini} & 4B & 4B & Apache 2.0 & general-purpose \\
\texttt{zen4} & 9B & 9B & Apache 2.0 & standard \\
\bottomrule
\end{longtable}

\subsection{Medium line (Qwen3.5 base, mixed dense/MoE)}

\begin{longtable}{p{3.5cm}rrlp{4cm}}
\toprule
\textbf{Model} & \textbf{Total} & \textbf{Active} & \textbf{License} & \textbf{Notes} \\
\midrule
\endhead
\texttt{zen4-pro} & 27B & 27B & Apache 2.0 & professional \\
\texttt{zen4-max} & 35B MoE & 3B & Apache 2.0 & MoE flagship (consumer) \\
\texttt{zen4-pro-max} & 80B MoE & 3B & Apache 2.0 & ultimate consumer \\
\texttt{zen4-mega} & 122B MoE & 10B & Apache 2.0 & higher-active MoE \\
\bottomrule
\end{longtable}

\subsection{Coder line}

\begin{longtable}{p{3.5cm}rrlp{4cm}}
\toprule
\textbf{Model} & \textbf{Total} & \textbf{Active} & \textbf{License} & \textbf{Notes} \\
\midrule
\endhead
\texttt{zen4-coder-flash} & 31B MoE & 3B & MIT & GLM-4.7-Flash base \\
\texttt{zen4-coder} & 80B MoE & 3B & Apache 2.0 & 512 experts \\
\texttt{zen4-coder-pro} & 355B & 355B & MIT & cloud-only \\
\bottomrule
\end{longtable}

\subsection{Frontier line}

\begin{longtable}{p{3.5cm}rrlp{4cm}}
\toprule
\textbf{Model} & \textbf{Total} & \textbf{Active} & \textbf{License} & \textbf{Notes} \\
\midrule
\endhead
\texttt{zen4-thunder} & 230B MoE & 10B & MIT & MiniMax-M2.5 base \\
\texttt{zen4-storm} & 456B MoE & 45B & MIT & MiniMax-M1 base \\
\texttt{zen4-titan} & 744B MoE & 40B & open & GLM-5 base \\
\texttt{zen4-ultra} & 1.04T MoE & 32B & open & Kimi K2.5 base \\
\texttt{zen4-ultra-gguf} & 1.04T MoE & 32B & open & K2.5 abliterated GGUF \\
\texttt{zen4-ultra-max} & 1T+ MoE & 50B & MIT & Ling-1T base \\
\bottomrule
\end{longtable}

\subsection{Vision-language sub-family}

The \texttt{zen-vl} line (released earlier and rolled into the per-LLM chain framework at activation) provides image+text models on Qwen3-VL bases at 4B, 8B, and 30B/MoE scales, each with an \texttt{-instruct} and \texttt{-agent} variant. Each VL model receives its own chain.

\subsection{Speech sub-family}

\texttt{zen-tts} (text-to-speech, multilingual, 24\,kHz output, voices: Cherry, Serena, Ethan, Chelsie, Momo) and \texttt{zen-omni} (unified text+vision+audio) each have their own chain. Modality-specific instructions extend the standard schema with audio-clip provenance records for \texttt{zen-tts} and per-modality data trees for \texttt{zen-omni}.

\subsection{Reranker, embedding, and specialty chains}

\begin{itemize}[nosep]
\item \texttt{zen-reranker} (ZIP-002) --- reranking model, dedicated chain.
\item \texttt{zen-embedding} --- embedding model.
\item \texttt{zen-guard}, \texttt{zen-guard-stream}, \texttt{zen-guard-gen} --- safety classifiers; share a single guard chain due to interlocking training.
\item \texttt{zen-scribe}, \texttt{zen-blog}, \texttt{zen-translator}, \texttt{zen-dub}, \texttt{zen-foley}, \texttt{zen-musician}, \texttt{zen-3d}, \texttt{zen-artist}, \texttt{zen-designer}, \texttt{zen-director}, \texttt{zen-video}, \texttt{zen-world}, \texttt{zen-voyager} --- specialty chains, each with its own training history and contributor graph.
\end{itemize}

\subsection{Lineage}

The base-model lineage table below records the upstream foundation each \Zen{} model derives from. The lineage is itself an artifact registered on the per-LLM chain (kind: \texttt{architecture}, with parent pointers to upstream sources where licensing permits). Brand policy: public docs do not name upstream organizations; the chain records them as cryptographic commitments only.

\begin{longtable}{p{4cm}p{8cm}}
\toprule
\textbf{\Zen{} model} & \textbf{Upstream foundation (chain-recorded)} \\
\midrule
\endhead
\texttt{zen4-nano} -- \texttt{zen4} & Qwen3.5 (Gated DeltaNet) \\
\texttt{zen4-pro} -- \texttt{zen4-mega} & Qwen3.5 abliterated lineage \\
\texttt{zen4-coder-flash} & GLM-4.7-Flash \\
\texttt{zen4-coder} & Qwen3-Coder-Next \\
\texttt{zen4-coder-pro} & GLM-4.7 \\
\texttt{zen4-thunder}, \texttt{zen4-storm} & MiniMax M-series \\
\texttt{zen4-titan} & GLM-5 \\
\texttt{zen4-ultra}, \texttt{zen4-ultra-gguf} & Kimi K2.5 \\
\texttt{zen4-ultra-max} & Ling-1T \\
\texttt{zen-vl-*} & Qwen3-VL \\
\texttt{zen-omni} & Qwen3-Omni \\
\bottomrule
\end{longtable}

\subsection{License and attribution}

Each \Zen{} chain inherits the upstream license (Apache 2.0 or MIT) and adds the chain-recorded contributor attribution graph as a binding revenue policy. Inference revenue routed from the \Hanzo{} AI Chain to the \Zen{} chain is distributed per-epoch according to the attribution tree at the time the inference was billed.