fix: remove 2T+ spec; position Zen 5 as largest open-source LLM ever

- zen5.mdx: replace '2T+ parameters' with 'The largest open-source LLM ever'
- zen5.mdx: update description + status banner to goal-first framing
- open-training post: same 2T+ → 'largest open-source LLM ever' fix
- homepage card: tease without parameter count
- zen-models: fix header comment Distilled → Diverse Experts

apps/zen-docs/app/(home)/page.tsx

@@ -132,7 +132,7 @@ export default function HomePage() {
             <div className="rounded-xl border border-fd-border bg-fd-background/60 p-4">
               <div className="text-xs font-semibold text-fd-primary mb-1">Zen MoDE Architecture</div>
               <p className="text-xs text-fd-muted-foreground leading-relaxed">
-                Mixture of Diverse Experts — 2T+ parameters, 2M+ context. Expert populations trained across diverse specializations for broad capability.
+                The largest open-source LLM ever built. Zen MoDE — Mixture of Diverse Experts — training now on Hanzo Network.
               </p>
             </div>
             <div className="rounded-xl border border-fd-border bg-fd-background/60 p-4">

apps/zen-docs/content/blog/open-training-hanzo-network.mdx

@@ -48,7 +48,7 @@ The result: anyone can verify that Zen 5 was trained on the compute and data we
 
 Zen 5 targets meaningful advances over Zen 4 across several dimensions:
 
-**Scale and architecture**: Zen MoDE (Mixture of Diverse Experts) at 2T+ total parameters. The key architectural innovation is expert diversity — experts are trained on different data mixtures with different inductive biases, and the router learns to leverage this diversity rather than just capacity. Per-token compute stays efficient via sparse activation.
+**Scale and architecture**: Zen MoDE (Mixture of Diverse Experts) — the largest open-source LLM ever built. The key architectural innovation is expert diversity: experts are trained on different data mixtures with different inductive biases, and the router learns to leverage this diversity rather than just capacity. Per-token compute stays efficient via sparse activation.
 
 **Context**: 2M+ token native context window. We believe multi-million-token context will become the baseline expectation for frontier models within two years. We are building for that now.
 

apps/zen-docs/content/docs/models/zen5.mdx

@@ -1,12 +1,12 @@
 ---
 title: Zen 5 — Research Preview
-description: Next-generation open model. Zen MoDE architecture. Training openly on Hanzo Network — contribute agentic data and help shape Zen 5.
+description: The largest open-source LLM ever built. Training openly on Hanzo Network — contribute agentic data and help shape Zen 5.
 ---
 
 # Zen 5 — Research Preview
 
 <div className="rounded-xl border border-fd-primary/30 bg-fd-primary/5 px-6 py-4 mb-8">
-  **Status:** In training · **Architecture:** Zen MoDE (Mixture of Diverse Experts) · **Open training:** [hanzo.network](https://hanzo.network)
+  **Status:** In training · **Goal:** Largest open-source LLM ever built · **Open training:** [hanzo.network](https://hanzo.network)
 </div>
 
 Zen 5 is the next generation of the Zen LM family. It is currently training openly on [Hanzo Network](https://hanzo.network) — decentralized compute with verifiable execution via NVIDIA Trusted Execution Environments (TEE).
@@ -18,7 +18,7 @@ We are accepting agentic training data submissions from the community. If you bu
 | Capability | Zen 4 | Zen 5 |
 |------------|-------|-------|
 | Architecture | Dense / standard MoE | Zen MoDE (Mixture of Diverse Experts) |
-| Scale | Up to 1T parameters | 2T+ parameters |
+| Scale | Up to 1T parameters | The largest open-source LLM ever |
 | Context | Up to 1M tokens | 2M+ tokens |
 | Training | Centralized | Open — Hanzo Network |
 | Reasoning | Chain-of-thought | Extended multi-step, tree-search |
@@ -33,7 +33,7 @@ We are accepting agentic training data submissions from the community. If you bu
 - **Expert diversity**: Each expert is initialized and fine-tuned on a different data mixture, creating genuine specialization rather than redundant capacity.
 - **Diversity-aware routing**: The router maximizes coverage across expert specializations rather than just capacity utilization.
 - **Cross-expert distillation**: Experts learn from each other during training — strong experts in one domain improve weak experts via targeted distillation steps.
-- **Sparse activation**: Only a small fraction of parameters activate per token, maintaining inference efficiency at 2T+ scale.
+- **Sparse activation**: Only a small fraction of parameters activate per token — frontier capability at a fraction of the inference cost.
 
 The result: a model that covers more ground at lower per-token compute than either dense models or standard MoE at equivalent parameter counts.
 

packages/zen-models/src/models.ts

@@ -4,8 +4,8 @@
  * Every Hanzo site (hanzo.ai, hanzo.industries, zenlm.org, etc.) imports from here.
  * Never duplicate model specs elsewhere.
  *
- * Zen MoDE (Mixture of Distilled Experts) — curating the best open-source
- * foundations and fusing them into a unified, high-performance family.
+ * Zen MoDE (Mixture of Diverse Experts) — Zen 5 architecture with deliberately
+ * diverse expert specializations for maximum coverage at efficient per-token compute.
  */
 
 import type { ZenModel } from './types'