diff --git a/studio/backend/core/inference/llama_cpp_builder.py b/studio/backend/core/inference/llama_cpp_builder.py
new file mode 100644
index 0000000000..268bddba8f
--- /dev/null
+++ b/studio/backend/core/inference/llama_cpp_builder.py
@@ -0,0 +1,248 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Background llama.cpp compilation.
+
+On server startup, if the llama-server binary is not found, a background
+thread clones and builds llama.cpp from source. Features that need
+llama.cpp (GGUF chat, export, AI Assist) can await `wait_for_ready()`
+which blocks until compilation finishes or fails.
+"""
+
+import os
+import shutil
+import subprocess
+import threading
+from pathlib import Path
+from typing import Optional
+
+import structlog
+from loggers import get_logger
+
+logger = get_logger(__name__)
+
+
+class LlamaCppBuilder:
+    """Manages background llama.cpp compilation."""
+
+    def __init__(self):
+        self._ready = threading.Event()
+        self._building = False
+        self._error: Optional[str] = None
+        self._thread: Optional[threading.Thread] = None
+        self._binary_path: Optional[str] = None
+
+    @property
+    def is_ready(self) -> bool:
+        return self._ready.is_set()
+
+    @property
+    def is_building(self) -> bool:
+        return self._building
+
+    @property
+    def error(self) -> Optional[str]:
+        return self._error
+
+    @property
+    def binary_path(self) -> Optional[str]:
+        return self._binary_path
+
+    def check_and_build(self) -> None:
+        """Check if llama-server exists. If not, start background build."""
+        from core.inference.llama_cpp import LlamaCppBackend
+
+        binary = LlamaCppBackend._find_llama_server_binary()
+        if binary:
+            self._binary_path = binary
+            self._ready.set()
+            logger.info(f"llama-server binary found: {binary}")
+            return
+
+        logger.info("llama-server binary not found, starting background build...")
+        self._building = True
+        self._thread = threading.Thread(
+            target=self._build, daemon=True, name="llama-cpp-build"
+        )
+        self._thread.start()
+
+    def wait_for_ready(self, timeout: Optional[float] = None) -> bool:
+        """Block until llama.cpp is ready. Returns True if ready, False on timeout."""
+        return self._ready.wait(timeout=timeout)
+
+    def _build(self) -> None:
+        """Clone and build llama.cpp in the background."""
+        try:
+            import sys
+
+            llama_dir = Path.home() / ".unsloth" / "llama.cpp"
+            binary_name = (
+                "llama-server.exe" if sys.platform == "win32" else "llama-server"
+            )
+
+            # Don't rebuild if binary appeared while we were starting
+            from core.inference.llama_cpp import LlamaCppBackend
+
+            binary = LlamaCppBackend._find_llama_server_binary()
+            if binary:
+                self._binary_path = binary
+                self._building = False
+                self._ready.set()
+                return
+
+            # Check prerequisites
+            if not shutil.which("cmake"):
+                self._error = "cmake not found. Install cmake to enable GGUF inference."
+                self._building = False
+                logger.warning(self._error)
+                return
+            if not shutil.which("git"):
+                self._error = "git not found. Install git to enable GGUF inference."
+                self._building = False
+                logger.warning(self._error)
+                return
+
+            # Clone
+            if llama_dir.exists():
+                shutil.rmtree(llama_dir, ignore_errors=True)
+
+            logger.info("Cloning llama.cpp...")
+            subprocess.run(
+                [
+                    "git",
+                    "clone",
+                    "--depth",
+                    "1",
+                    "https://github.com/ggml-org/llama.cpp.git",
+                    str(llama_dir),
+                ],
+                check=True,
+                capture_output=True,
+            )
+
+            # Detect CUDA
+            cmake_args = [
+                "-DBUILD_SHARED_LIBS=OFF",
+            ]
+            nvcc = shutil.which("nvcc")
+            if not nvcc:
+                for cuda_dir in sorted(
+                    Path("/usr/local").glob("cuda-*/bin/nvcc"), reverse=True
+                ):
+                    nvcc = str(cuda_dir)
+                    break
+            if not nvcc and Path("/usr/local/cuda/bin/nvcc").is_file():
+                nvcc = "/usr/local/cuda/bin/nvcc"
+
+            if nvcc:
+                logger.info(f"CUDA detected: {nvcc}")
+                cmake_args.append("-DGGML_CUDA=ON")
+                # Detect compute capabilities
+                try:
+                    result = subprocess.run(
+                        [
+                            "nvidia-smi",
+                            "--query-gpu=compute_cap",
+                            "--format=csv,noheader",
+                        ],
+                        capture_output=True,
+                        text=True,
+                        timeout=10,
+                    )
+                    if result.returncode == 0:
+                        caps = set()
+                        for line in result.stdout.strip().split("\n"):
+                            cap = line.strip().replace(".", "")
+                            if cap:
+                                caps.add(cap)
+                        if caps:
+                            cmake_args.append(
+                                f"-DCMAKE_CUDA_ARCHITECTURES={';'.join(sorted(caps))}"
+                            )
+                except Exception:
+                    pass
+                cmake_args.append("-DCMAKE_CUDA_FLAGS=--threads=0")
+            else:
+                logger.info("No CUDA detected, building CPU-only llama.cpp")
+
+            # Use Ninja if available
+            generator_args = []
+            if shutil.which("ninja"):
+                generator_args = ["-G", "Ninja"]
+
+            # Configure
+            build_dir = llama_dir / "build"
+            logger.info("Configuring llama.cpp build...")
+            subprocess.run(
+                [
+                    "cmake",
+                    "-B",
+                    str(build_dir),
+                    "-S",
+                    str(llama_dir),
+                ]
+                + generator_args
+                + cmake_args,
+                check=True,
+                capture_output=True,
+            )
+
+            # Build
+            logger.info("Building llama.cpp (this may take a few minutes)...")
+            subprocess.run(
+                [
+                    "cmake",
+                    "--build",
+                    str(build_dir),
+                    "--config",
+                    "Release",
+                    "--target",
+                    "llama-server",
+                    "llama-quantize",
+                    "-j",
+                ],
+                check=True,
+                capture_output=True,
+            )
+
+            # Symlink llama-quantize for unsloth-zoo's check_llama_cpp()
+            quantize_src = build_dir / "bin" / "llama-quantize"
+            quantize_link = llama_dir / "llama-quantize"
+            if quantize_src.is_file() and not quantize_link.exists():
+                try:
+                    quantize_link.symlink_to(quantize_src)
+                except Exception:
+                    pass
+
+            # Verify
+            expected = build_dir / "bin" / binary_name
+            if expected.is_file():
+                self._binary_path = str(expected)
+                logger.info(f"llama.cpp build complete: {self._binary_path}")
+            else:
+                self._error = f"Build completed but binary not found at {expected}"
+                logger.error(self._error)
+
+        except subprocess.CalledProcessError as e:
+            stderr = e.stderr.decode() if e.stderr else ""
+            self._error = f"llama.cpp build failed: {stderr[-500:]}"
+            logger.error(self._error)
+        except Exception as e:
+            self._error = f"llama.cpp build error: {e}"
+            logger.error(self._error)
+        finally:
+            self._building = False
+            self._ready.set()  # Unblock waiters even on failure
+
+
+# ── Singleton ────────────────────────────────────────────────────
+
+_builder: Optional[LlamaCppBuilder] = None
+
+
+def get_llama_cpp_builder() -> LlamaCppBuilder:
+    global _builder
+    if _builder is None:
+        _builder = LlamaCppBuilder()
+    return _builder
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index ecc11c30d5..af20146a7f 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -2352,11 +2352,12 @@ class UnslothTrainer:
                     logger.info("Stopped during dataset loading\n")
                     return None
 
+                n_rows = len(dataset) if hasattr(dataset, "__len__") else 0
                 self._update_progress(
-                    status_message = f"Loaded dataset from HuggingFace: {dataset_source}"
+                    status_message = f"Loaded dataset from HuggingFace: {dataset_source} ({n_rows:,} rows)"
                 )
                 logger.info(
-                    f"Loaded dataset from Hugging Face: {dataset_source} ({len(dataset)} rows)\n"
+                    f"Loaded dataset from Hugging Face: {dataset_source} ({n_rows:,} rows)\n"
                 )
 
                 # Resolve eval split from a separate HF split (explicit or auto-detected)
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 3ab846306e..4495c99e2f 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -75,6 +75,12 @@ async def lifespan(app: FastAPI):
 
     threading.Thread(target = _precache, daemon = True).start()
 
+    # Build llama.cpp in the background if the binary is missing.
+    # Studio starts immediately; GGUF features wait for the build to finish.
+    from core.inference.llama_cpp_builder import get_llama_cpp_builder
+
+    get_llama_cpp_builder().check_and_build()
+
     if storage.ensure_default_admin():
         bootstrap_pw = storage.get_bootstrap_password()
         app.state.bootstrap_password = bootstrap_pw
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index e93ba1d6d6..34a5f345cb 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -111,6 +111,19 @@ async def load_model(
 
         # ── GGUF path: load via llama-server ──────────────────────
         if config.is_gguf:
+            # Wait for llama.cpp to be compiled if a background build is in progress
+            from core.inference.llama_cpp_builder import get_llama_cpp_builder
+
+            builder = get_llama_cpp_builder()
+            if not builder.is_ready:
+                logger.info("Waiting for llama.cpp build to finish...")
+                ready = await asyncio.to_thread(builder.wait_for_ready, timeout = 600)
+                if not ready or builder.error:
+                    raise HTTPException(
+                        status_code = 503,
+                        detail = builder.error or "llama.cpp is still compiling. Please wait.",
+                    )
+
             llama_backend = get_llama_cpp_backend()
             unsloth_backend = get_inference_backend()
 
@@ -538,6 +551,19 @@ async def get_status(
         raise HTTPException(status_code = 500, detail = f"Failed to get status: {str(e)}")
 
 
+@router.get("/llama-cpp-status")
+async def llama_cpp_status():
+    """Check if llama.cpp is ready (built and available)."""
+    from core.inference.llama_cpp_builder import get_llama_cpp_builder
+
+    builder = get_llama_cpp_builder()
+    return {
+        "ready": builder.is_ready and builder.error is None,
+        "building": builder.is_building,
+        "error": builder.error,
+    }
+
+
 # =====================================================================
 # Audio (TTS) Generation  (/audio/generate)
 # =====================================================================
diff --git a/studio/backend/utils/datasets/dataset_utils.py b/studio/backend/utils/datasets/dataset_utils.py
index 9d15b86ca1..1d2e4083ec 100644
--- a/studio/backend/utils/datasets/dataset_utils.py
+++ b/studio/backend/utils/datasets/dataset_utils.py
@@ -1092,6 +1092,11 @@ def format_and_template_dataset(
     # LLM FLOW (Existing code)
     else:
         # Step 1: Format the dataset
+        n_rows = len(dataset) if hasattr(dataset, "__len__") else None
+        if progress_callback and n_rows:
+            progress_callback(
+                status_message = f"Formatting dataset ({n_rows:,} rows)..."
+            )
         dataset_info = format_dataset(
             dataset,
             format_type = format_type,
@@ -1106,6 +1111,10 @@ def format_and_template_dataset(
         )
 
         # Step 2: Apply chat template
+        if progress_callback and n_rows:
+            progress_callback(
+                status_message = f"Applying chat template ({n_rows:,} rows)..."
+            )
         # Gemma emits a leading <bos> that must be stripped for text-only chatml/sharegpt.
         is_alpaca = format_type == "alpaca" or (
             format_type == "auto" and dataset_info["detected_format"] == "alpaca"
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 01bc762d86..0f63e0d462 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -62,6 +62,15 @@ export async function getInferenceStatus(): Promise<InferenceStatusResponse> {
   return parseJsonOrThrow<InferenceStatusResponse>(response);
 }
 
+export async function getLlamaCppStatus(): Promise<{
+  ready: boolean;
+  building: boolean;
+  error: string | null;
+}> {
+  const response = await authFetch("/api/inference/llama-cpp-status");
+  return parseJsonOrThrow(response);
+}
+
 export async function loadModel(
   payload: LoadModelRequest,
 ): Promise<LoadModelResponse> {
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index b79c3073b2..e06418ce6d 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -9,6 +9,7 @@ import {
   getDownloadProgress,
   getGgufDownloadProgress,
   getInferenceStatus,
+  getLlamaCppStatus,
   listLoras,
   listModels,
   loadModel,
@@ -318,6 +319,25 @@ export function useChatModelRuntime() {
       try {
         async function performLoad(): Promise<void> {
           if (abortCtrl.signal.aborted) throw new Error("Cancelled");
+
+          // For GGUF models, check if llama.cpp is still compiling
+          if (ggufVariant) {
+            try {
+              const llamaStatus = await getLlamaCppStatus();
+              if (llamaStatus.building) {
+                toast.info("Waiting for llama.cpp to compile...", {
+                  description: "This is a one-time build. Studio is still usable for non-GGUF tasks.",
+                  duration: 8000,
+                });
+              } else if (llamaStatus.error) {
+                throw new Error(llamaStatus.error);
+              }
+            } catch (e) {
+              if (e instanceof Error && e.message.includes("not found")) throw e;
+              // Endpoint might not exist on older backends, ignore
+            }
+          }
+
           let previousWasUnloaded = false;
           const currentCheckpoint =
             useChatRuntimeStore.getState().params.checkpoint;