diff --git a/.claude/docs/working/EXERCISE_ENRICHMENT_ADMIN.md b/.claude/docs/working/EXERCISE_ENRICHMENT_ADMIN.md
index 16cf56a..8d00e20 100644
--- a/.claude/docs/working/EXERCISE_ENRICHMENT_ADMIN.md
+++ b/.claude/docs/working/EXERCISE_ENRICHMENT_ADMIN.md
@@ -40,7 +40,9 @@ Wiederverwendet `run_exercise_form_ai_suggestion` → Prompts `exercise_skill_su
 - Kandidaten: **Status** primär (Default `draft`); Sichtbarkeit Default **`private`**, wählbar bis „Alle“
 - Skill-Merge Default: **`replace_all`** (alle Skills KI-neu, `ai_suggested=true` — unterscheidbar von manuell)
 - Nach Apply: `set_status=in_review` (nie automatisch `approved`)
-- Batch: keine Gesamtgrenze (bis 10.000 IDs); **Analyze** + explizite Nutzerbestätigung; HTTP-Chunks à 25/100
+- Batch: keine Gesamtgrenze (bis 10.000 IDs); **Analyze** + explizite Nutzerbestätigung
+- **Preview:** max. **3 Übungen/HTTP-Request** (parallel LLM), Frontend chunked — vermeidet Gateway-504 (~60s Fritz!Box)
+- **Apply:** HTTP-Chunks à 25 (nur DB, kein LLM)
 
 ## Inhalte (modular)
 
diff --git a/backend/exercise_enrichment.py b/backend/exercise_enrichment.py
index f9da6c6..659226a 100644
--- a/backend/exercise_enrichment.py
+++ b/backend/exercise_enrichment.py
@@ -22,8 +22,10 @@ SkillMergeMode = Literal["additive", "replace_ai_only", "replace_all"]
 
 SKILL_MERGE_MODES = frozenset({"additive", "replace_ai_only", "replace_all"})
 DEFAULT_SET_STATUS = "in_review"
-# Max. IDs pro einzelner HTTP-Anfrage (Frontend chunked darüber hinaus).
-MAX_BATCH_EXERCISES = 100
+# Max. IDs pro Apply-HTTP-Anfrage (kein LLM).
+MAX_BATCH_EXERCISES = 50
+# Preview: pro Request nur wenige Übungen — sonst Gateway-504 (Fritz!Box o.ä. ~60s).
+MAX_PREVIEW_BATCH_EXERCISES = 3
 
 _INSTRUCTION_FIELDS = ("goal", "execution", "preparation", "trainer_notes")
 _SKILL_COMPARE_KEYS = ("intensity", "required_level", "target_level", "is_primary")
diff --git a/backend/routers/exercise_enrichment_admin.py b/backend/routers/exercise_enrichment_admin.py
index 2f6a248..5b370f4 100644
--- a/backend/routers/exercise_enrichment_admin.py
+++ b/backend/routers/exercise_enrichment_admin.py
@@ -6,6 +6,7 @@ Siehe ACCESS_LAYER_ENDPOINT_AUDIT.md.
 """
 from __future__ import annotations
 
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from typing import Any, Dict, List, Literal, Optional
 
 from fastapi import APIRouter, Depends, HTTPException, Query
@@ -17,6 +18,7 @@ from db import get_cursor, get_db, r2d
 from exercise_enrichment import (
     DEFAULT_SET_STATUS,
     MAX_BATCH_EXERCISES,
+    MAX_PREVIEW_BATCH_EXERCISES,
     SKILL_MERGE_MODES,
     SkillMergeMode,
     apply_exercise_enrichment,
@@ -24,6 +26,8 @@ from exercise_enrichment import (
     preview_exercise_enrichment,
 )
 
+_PREVIEW_MAX_WORKERS = 3
+
 router = APIRouter(tags=["admin_exercise_enrichment"])
 
 _VALID_STATUS_FILTER = frozenset({"draft", "in_review", "approved", "archived"})
@@ -108,7 +112,7 @@ class EnrichmentModes(BaseModel):
 
 
 class EnrichmentPreviewBody(BaseModel):
-    exercise_ids: List[int] = Field(..., min_length=1, max_length=MAX_BATCH_EXERCISES)
+    exercise_ids: List[int] = Field(..., min_length=1, max_length=MAX_PREVIEW_BATCH_EXERCISES)
     modes: EnrichmentModes = Field(default_factory=EnrichmentModes)
     merge_mode: SkillMergeMode = "replace_all"
 
@@ -301,12 +305,40 @@ def analyze_enrichment(body: EnrichmentAnalyzeBody, session: dict = Depends(requ
     }
 
 
+def _preview_one_exercise(
+    ex_id: int,
+    *,
+    want_skills: bool,
+    want_summary: bool,
+    want_instructions: bool,
+    merge_mode: SkillMergeMode,
+) -> Dict[str, Any]:
+    """Einzel-Preview mit eigener DB-Connection (Thread-Pool)."""
+    try:
+        with get_db() as conn:
+            cur = get_cursor(conn)
+            row = preview_exercise_enrichment(
+                cur,
+                ex_id,
+                want_skills=want_skills,
+                want_summary=want_summary,
+                want_instructions=want_instructions,
+                merge_mode=merge_mode,
+            )
+        return row
+    except HTTPException as he:
+        d = he.detail
+        return {"exercise_id": ex_id, "ok": False, "error": d if isinstance(d, str) else str(d)}
+    except Exception as exc:  # pragma: no cover
+        return {"exercise_id": ex_id, "ok": False, "error": str(exc)}
+
+
 @router.post("/api/admin/exercise-enrichment/preview")
 def preview_enrichment(body: EnrichmentPreviewBody, session: dict = Depends(require_auth)):
-    """Dry-Run: KI-Vorschläge laden, nichts speichern."""
+    """Dry-Run: KI-Vorschläge laden, nichts speichern (max. 3 Übungen/Request, parallel)."""
     _require_superadmin(session)
 
-    ids = _normalize_id_list(body.exercise_ids, max_items=MAX_BATCH_EXERCISES)
+    ids = _normalize_id_list(body.exercise_ids, max_items=MAX_PREVIEW_BATCH_EXERCISES)
     if not ids:
         raise HTTPException(status_code=400, detail="Keine gültigen Übungs-IDs")
 
@@ -320,27 +352,28 @@ def preview_enrichment(body: EnrichmentPreviewBody, session: dict = Depends(requ
     results: List[Dict[str, Any]] = []
     errors: List[Dict[str, Any]] = []
 
-    with get_db() as conn:
-        cur = get_cursor(conn)
-        for ex_id in ids:
-            try:
-                row = preview_exercise_enrichment(
-                    cur,
-                    ex_id,
-                    want_skills=modes.skills,
-                    want_summary=modes.summary,
-                    want_instructions=modes.instructions,
-                    merge_mode=body.merge_mode,
-                )
-                if row.get("ok"):
-                    results.append(row)
-                else:
-                    errors.append(row)
-            except HTTPException as he:
-                d = he.detail
-                errors.append({"exercise_id": ex_id, "ok": False, "error": d if isinstance(d, str) else str(d)})
-            except Exception as exc:  # pragma: no cover
-                errors.append({"exercise_id": ex_id, "ok": False, "error": str(exc)})
+    workers = min(_PREVIEW_MAX_WORKERS, len(ids))
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        futures = [
+            pool.submit(
+                _preview_one_exercise,
+                ex_id,
+                want_skills=modes.skills,
+                want_summary=modes.summary,
+                want_instructions=modes.instructions,
+                merge_mode=body.merge_mode,
+            )
+            for ex_id in ids
+        ]
+        for fut in as_completed(futures):
+            row = fut.result()
+            if row.get("ok"):
+                results.append(row)
+            else:
+                errors.append(row)
+
+    results.sort(key=lambda r: int(r.get("exercise_id") or 0))
+    errors.sort(key=lambda r: int(r.get("exercise_id") or 0))
 
     est = estimate_llm_calls(
         exercise_count=len(ids),
diff --git a/backend/version.py b/backend/version.py
index a1a705d..5130641 100644
--- a/backend/version.py
+++ b/backend/version.py
@@ -1,6 +1,6 @@
 # Shinkan Jinkendo Version Information
 
-APP_VERSION = "0.8.179"
+APP_VERSION = "0.8.180"
 BUILD_DATE = "2026-05-23"
 DB_SCHEMA_VERSION = "20260531074"
 
@@ -19,7 +19,7 @@ MODULE_VERSIONS = {
     "media_legal_hold": "1.0.0",  # P-11: Sofortsperre-Services (set_legal_hold, release_legal_hold)
     "media_lifecycle": "1.1.0",  # P-11: Retention-Job ueberspringt Legal-Hold-Assets
     "admin_ai_skill_retrieval": "1.0.0",  # Superadmin CRUD /api/admin/ai-skill-retrieval-profiles (Migration 068)
-    "exercise_enrichment_admin": "1.1.0",  # Analyze, candidate-ids, instructions/summary apply; unbegrenzte Batch-Auswahl
+    "exercise_enrichment_admin": "1.1.1",  # Preview max 3/Request + parallel LLM (Gateway-504 vermeiden)
     "admin_ai_prompts": "1.0.3",  # Migration 070: openrouter_model; PUT/Liste/Detail
     "ai_prompt_job": "0.2.1",  # want_instructions; run_exercise_form_ai_suggestion
     "ai_prompt_context": "0.2.0",  # preparation/trainer_notes; has_instruction_source_text
@@ -44,6 +44,13 @@ MODULE_VERSIONS = {
 }
 
 CHANGELOG = [
+    {
+        "version": "0.8.180",
+        "date": "2026-05-23",
+        "changes": [
+            "Fix: Übungs-Anreicherung Vorschau — max. 3 Übungen/Request, parallel LLM; Frontend-Pakete + Retry bei 504.",
+        ],
+    },
     {
         "version": "0.8.179",
         "date": "2026-05-23",
diff --git a/frontend/src/pages/AdminExerciseEnrichmentPage.jsx b/frontend/src/pages/AdminExerciseEnrichmentPage.jsx
index a31b431..e9cd6c7 100644
--- a/frontend/src/pages/AdminExerciseEnrichmentPage.jsx
+++ b/frontend/src/pages/AdminExerciseEnrichmentPage.jsx
@@ -43,7 +43,9 @@ const INSTRUCTION_LABELS = {
   trainer_notes: 'Trainer-Hinweise',
 }
 
-const CHUNK_SIZE = 25
+/** Preview: kleine Pakete — Gateway (Fritz!Box o.ä.) timeout oft ~60s bei LLM-Ketten. */
+const PREVIEW_CHUNK_SIZE = 3
+const APPLY_CHUNK_SIZE = 25
 
 function skillLabel(sk) {
   if (!sk) return '—'
@@ -177,9 +179,15 @@ function RunDialog({
               )}
             </div>
           )}
+          {exerciseCount >= 10 && (
+            <p style={{ margin: '10px 0 0', color: 'var(--text2)', fontSize: '0.88rem' }}>
+              Die Vorschau läuft in Paketen à 3 Übungen (Gateway-Timeout vermeiden). Fenster offen lassen —
+              bei {exerciseCount} Übungen ca. {Math.ceil(exerciseCount / 3)} Pakete.
+            </p>
+          )}
           {exerciseCount >= 25 && (
-            <p style={{ margin: '10px 0 0', color: 'var(--danger)', fontSize: '0.88rem' }}>
-              Großer Batch — OpenRouter-Kosten und Laufzeit beachten. Der Lauf kann mehrere Minuten dauern.
+            <p style={{ margin: '6px 0 0', color: 'var(--danger)', fontSize: '0.88rem' }}>
+              Großer Batch — OpenRouter-Kosten und Gesamtlaufzeit (mehrere Minuten) beachten.
             </p>
           )}
         </div>
@@ -432,6 +440,24 @@ export default function AdminExerciseEnrichmentPage() {
     setDialogOpen(true)
   }
 
+  async function previewChunkWithRetry(chunk, attempt = 0) {
+    try {
+      return await api.previewExerciseEnrichment({
+        exercise_ids: chunk,
+        modes,
+        merge_mode: mergeMode,
+      })
+    } catch (e) {
+      const msg = e.message || String(e)
+      const isTimeout = /504|502|timeout/i.test(msg)
+      if (isTimeout && attempt < 2) {
+        await new Promise((r) => setTimeout(r, 2500))
+        return previewChunkWithRetry(chunk, attempt + 1)
+      }
+      throw e
+    }
+  }
+
   async function runPreviewFromDialog() {
     setDialogOpen(false)
     setError('')
@@ -440,18 +466,21 @@ export default function AdminExerciseEnrichmentPage() {
     const allResults = []
     const allErrors = []
     let estTotal = 0
+    const totalChunks = Math.ceil(selectedIds.length / PREVIEW_CHUNK_SIZE)
 
     setJobProgress({ done: 0, total: selectedIds.length, phase: 'Vorschau' })
 
     try {
-      for (let i = 0; i < selectedIds.length; i += CHUNK_SIZE) {
+      for (let i = 0; i < selectedIds.length; i += PREVIEW_CHUNK_SIZE) {
         if (abortRef.current) break
-        const chunk = selectedIds.slice(i, i + CHUNK_SIZE)
-        const resp = await api.previewExerciseEnrichment({
-          exercise_ids: chunk,
-          modes,
-          merge_mode: mergeMode,
+        const chunk = selectedIds.slice(i, i + PREVIEW_CHUNK_SIZE)
+        const chunkNo = Math.floor(i / PREVIEW_CHUNK_SIZE) + 1
+        setJobProgress({
+          done: i,
+          total: selectedIds.length,
+          phase: `Vorschau Paket ${chunkNo}/${totalChunks}`,
         })
+        const resp = await previewChunkWithRetry(chunk)
         allResults.push(...(resp.results || []))
         allErrors.push(...(resp.errors || []))
         const est = resp.estimated_llm_calls
@@ -459,7 +488,7 @@ export default function AdminExerciseEnrichmentPage() {
         setJobProgress({
           done: Math.min(i + chunk.length, selectedIds.length),
           total: selectedIds.length,
-          phase: 'Vorschau',
+          phase: `Vorschau Paket ${chunkNo}/${totalChunks}`,
         })
       }
       setPreviewRows(allResults)
@@ -517,9 +546,9 @@ export default function AdminExerciseEnrichmentPage() {
     setJobProgress({ done: 0, total: applyItems.length, phase: 'Anwenden' })
 
     try {
-      for (let i = 0; i < applyItems.length; i += CHUNK_SIZE) {
+      for (let i = 0; i < applyItems.length; i += APPLY_CHUNK_SIZE) {
         if (abortRef.current) break
-        const chunk = applyItems.slice(i, i + CHUNK_SIZE)
+        const chunk = applyItems.slice(i, i + APPLY_CHUNK_SIZE)
         const resp = await api.applyExerciseEnrichment({
           items: chunk,
           modes: previewMeta?.modes || modes,