diff --git a/.github/workflows/embeddings.yml b/.github/workflows/embeddings.yml
new file mode 100644
index 0000000000000..a2aa614a2aa76
--- /dev/null
+++ b/.github/workflows/embeddings.yml
@@ -0,0 +1,59 @@
+# Embedding CLI build and tests
+name: Embedding CLI
+
+on:
+    workflow_dispatch:
+    push:
+        branches:
+            - feature/*
+            - master
+        paths:
+            - '.github/workflows/embeddings.yml'
+            - 'examples/embedding/**'
+            - 'examples/tests/**'
+    pull_request:
+        types: [opened, synchronize, reopened]
+        paths:
+            - '.github/workflows/embeddings.yml'
+            - 'examples/embedding/**'
+            - 'examples/tests/**'
+
+jobs:
+    embedding-cli-tests:
+        runs-on: ubuntu-latest
+
+        steps:
+            - name: Install system deps
+              run: |
+                  sudo apt-get update
+                  sudo apt-get -y install \
+                    build-essential \
+                    cmake \
+                    curl \
+                    python3-pip \
+                    libcurl4-openssl-dev
+
+            - name: Checkout repository
+              uses: actions/checkout@v4
+              with:
+                  fetch-depth: 0
+
+            - name: Set up Python
+              uses: actions/setup-python@v5
+              with:
+                  python-version: '3.11'
+
+            - name: Install Python deps
+              run: |
+                  pip install -r requirements.txt || echo "No extra requirements found"
+                  pip install pytest
+
+            - name: Build llama-embedding
+              run: |
+                  cmake -B build \
+                    -DCMAKE_BUILD_TYPE=Release
+                  cmake --build build --target llama-embedding -j $(nproc)
+
+            - name: Run embedding tests
+              run: |
+                  pytest -v examples/tests
diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp
index 9e3ab5905bb37..29dd5a8615ec9 100644
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -81,13 +81,11 @@ static void print_raw_embeddings(const float * emb,
     const bool is_rank = (pooling_type == LLAMA_POOLING_TYPE_RANK);
     const int cols = is_rank ? std::min<int>(n_embd, (int) n_cls_out) : n_embd;
 
+    const char *fmt = embd_normalize == 0 ? "%1.0f%s" : "%1.7f%s";
+
     for (int j = 0; j < n_embd_count; ++j) {
         for (int i = 0; i < cols; ++i) {
-            if (embd_normalize == 0) {
-                LOG("%1.0f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
-            } else {
-                LOG("%1.7f%s", emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
-            }
+            LOG(fmt, emb[j * n_embd + i], (i + 1 < cols ? " " : ""));
         }
         LOG("\n");
     }
diff --git a/examples/tests/__init__.py b/examples/tests/__init__.py
new file mode 100644
index 0000000000000..e69de29bb2d1d
diff --git a/examples/tests/test_embedding.py b/examples/tests/test_embedding.py
new file mode 100644
index 0000000000000..72dfdabe4600a
--- /dev/null
+++ b/examples/tests/test_embedding.py
@@ -0,0 +1,291 @@
+from concurrent.futures import ProcessPoolExecutor, as_completed
+from functools import wraps
+import numpy as np
+from pathlib import Path
+import json, os, time, statistics, subprocess, math
+
+
+# ---------------------------------------------------------------------------
+# Benchmark decorator
+# ---------------------------------------------------------------------------
+
+def benchmark(n=3):
+    def decorator(fn):
+        @wraps(fn)
+        def wrapper(*args, **kwargs):
+            times = []
+            result = None
+            for _ in range(n):
+                start = time.perf_counter()
+                result = fn(*args, **kwargs)
+                times.append(time.perf_counter() - start)
+            avg = statistics.mean(times)
+            print(f"\n[benchmark] {fn.__name__}: mean={avg*1000:.1f} ms over {n} runs")
+            return result
+        return wrapper
+    return decorator
+
+
+# ---------------------------------------------------------------------------
+# Model helpers
+# ---------------------------------------------------------------------------
+
+def get_model_hf_params():
+    """Default lightweight embedding model."""
+    return {
+        "hf_repo": "ggml-org/embeddinggemma-300M-qat-q4_0-GGUF",
+        "hf_file": "embeddinggemma-300M-qat-Q4_0.gguf",
+    }
+
+
+def ensure_model_downloaded(params=None):
+    repo_root = Path(__file__).resolve().parents[2]
+    cache_dir = os.environ.get("LLAMA_CACHE", "tmp")
+    emb_path = repo_root / "build/bin/llama-embedding"
+    if not emb_path.exists() and os.name == "nt":
+        emb_path = repo_root / "build/bin/Release/llama-embedding.exe"
+    if not emb_path.exists():
+        raise FileNotFoundError(f"llama-embedding not found at {emb_path}")
+
+    params = params or get_model_hf_params()
+    cmd = [
+        str(emb_path),
+        "-hfr", params["hf_repo"],
+        "-hff", params["hf_file"],
+        "--ctx-size", "16",
+        "--embd-output-format", "json",
+        "--no-warmup",
+        "--threads", "1",
+    ]
+
+    env = os.environ.copy()
+    env["LLAMA_CACHE"] = cache_dir
+    result = subprocess.run(cmd, input="ok", capture_output=True, text=True, env=env)
+    if result.returncode != 0:
+        raise RuntimeError(f"Failed to download model:\n{result.stderr}")
+    return params
+
+
+def run_embedding(text: str, fmt: str = "raw", params=None):
+    repo_root = Path(__file__).resolve().parents[2]
+    exe = repo_root / "build/bin/llama-embedding"
+    assert exe.exists(), f"Missing binary: {exe}"
+
+    params = ensure_model_downloaded(params)
+    cache_dir = os.environ.get("LLAMA_CACHE", "tmp")
+
+    cmd = [
+        str(exe),
+        "-hfr", params["hf_repo"],
+        "-hff", params["hf_file"],
+        "--ctx-size", "2048",
+        "--embd-output-format", fmt,
+    ]
+
+    env = os.environ.copy()
+    env["LLAMA_CACHE"] = cache_dir
+
+    out = subprocess.run(cmd, input=text, capture_output=True, text=True, env=env)
+    if out.returncode != 0:
+        print(out.stderr)
+        raise AssertionError(f"embedding binary failed (code {out.returncode})")
+    return out.stdout.strip()
+
+
+# ---------------------------------------------------------------------------
+# 1️⃣ RAW vs JSON baseline tests
+# ---------------------------------------------------------------------------
+
+@benchmark(n=3)
+def test_embedding_raw_and_json_consistency():
+    """
+    Run both output modes and verify same embedding shape, norm similarity,
+    and small cosine distance.
+    """
+    out_raw = run_embedding("hello world", "raw")
+    floats_raw = np.array([float(x) for x in out_raw.split()])
+
+    out_json = run_embedding("hello world", "json")
+    j = json.loads(out_json)
+    floats_json = np.array(j["data"][0]["embedding"])
+
+    assert len(floats_raw) == len(floats_json), "Embedding dimension mismatch"
+    cos = np.dot(floats_raw, floats_json) / (np.linalg.norm(floats_raw) * np.linalg.norm(floats_json))
+    print(f"Cosine similarity raw vs json: {cos:.4f}")
+    # expect high similarity but not perfect (formatting precision differences)
+    assert cos > 0.999, f"Unexpected divergence between raw and json output ({cos:.4f})"
+
+
+@benchmark(n=3)
+def test_embedding_perf_regression_raw_vs_json():
+    """
+    Compare performance between raw and json output.
+    Ensures raw mode is not significantly slower or memory-heavier.
+    """
+    text = "performance regression test " * 512
+    params = ensure_model_downloaded()
+
+    def run(fmt):
+        start = time.perf_counter()
+        out = run_embedding(text, fmt, params)
+        dur = time.perf_counter() - start
+        mem = len(out)
+        return dur, mem
+
+    t_raw, m_raw = run("raw")
+    t_json, m_json = run("json")
+
+    print(f"[perf] raw={t_raw:.3f}s ({m_raw/1e3:.1f} KB) | json={t_json:.3f}s ({m_json/1e3:.1f} KB)")
+    # raw should never be significantly slower or consume wildly more memory
+    assert t_raw <= t_json * 1.2, f"raw too slow vs json ({t_raw:.3f}s vs {t_json:.3f}s)"
+    assert m_raw <= m_json * 1.2, f"raw output unexpectedly larger ({m_raw} vs {m_json} bytes)"
+
+
+# ---------------------------------------------------------------------------
+# 2️⃣ Edge-case coverage
+# ---------------------------------------------------------------------------
+
+def test_embedding_empty_input():
+    """
+    Empty input should not crash and should yield a deterministic, finite embedding.
+    Some models (e.g. Gemma/BGE) emit BOS token embedding with norm ≈ 1.0.
+    """
+    out1 = run_embedding("", "raw")
+    out2 = run_embedding("", "raw")
+
+    floats1 = np.array([float(x) for x in out1.split()])
+    floats2 = np.array([float(x) for x in out2.split()])
+
+    # Basic validity
+    assert len(floats1) > 0, "Empty input produced no embedding"
+    assert np.all(np.isfinite(floats1)), "Embedding contains NaN or inf"
+    norm = np.linalg.norm(floats1)
+    assert 0.5 <= norm <= 1.5, f"Unexpected norm for empty input: {norm}"
+
+    # Determinism check: cosine similarity should be ≈ 1
+    cos = np.dot(floats1, floats2) / (np.linalg.norm(floats1) * np.linalg.norm(floats2))
+    assert cos > 0.9999, f"Empty input not deterministic (cos={cos:.4f})"
+    print(f"[empty] norm={norm:.4f}, cos={cos:.6f}")
+
+
+def test_embedding_special_characters():
+    """Unicode and punctuation coverage."""
+    special_text = "你好 🌍\n\t!@#$%^&*()_+-=[]{}|;:'\",.<>?/`~"
+    out = run_embedding(special_text, "raw")
+    floats = [float(x) for x in out.split()]
+    assert len(floats) > 10
+    norm = np.linalg.norm(floats)
+    assert math.isfinite(norm) and norm > 0
+
+
+@benchmark(n=1)
+def test_embedding_very_long_input():
+    """Stress test for context limit handling."""
+    long_text = "lorem " * 10000
+    out = run_embedding(long_text, "raw")
+    floats = [float(x) for x in out.split()]
+    print(f"Output floats (long input): {len(floats)}")
+    assert len(floats) > 100
+    assert np.isfinite(np.linalg.norm(floats))
+
+
+# ---------------------------------------------------------------------------
+# 3️⃣ Legacy and concurrency coverage (unchanged)
+# ---------------------------------------------------------------------------
+
+@benchmark(n=3)
+def test_embedding_raw_vector_shape():
+    out = run_embedding("hello world", "raw")
+    floats = [float(x) for x in out.split()]
+    print(f"Embedding size: {len(floats)} floats")
+    assert len(floats) > 100
+    norm = np.linalg.norm(floats)
+    assert 0.5 < norm < 2.0
+
+
+@benchmark(n=3)
+def test_embedding_large_vector_output():
+    text = " ".join(["hello"] * 4096)
+    out = run_embedding(text, "raw")
+    valid_dims = {384, 768, 1024, 1280, 2048, 4096}
+    floats = [float(x) for x in out.split()]
+    print(f"Output floats: {len(floats)}")
+    assert len(floats) in valid_dims, (
+        f"Unexpected embedding size: {len(floats)}. Expected one of {sorted(valid_dims)}."
+    )
+
+
+def run_one(args):
+    i, params, text = args
+    repo_root = Path(__file__).resolve().parents[2]
+    exe = repo_root / "build/bin/llama-embedding"
+    cache_dir = os.environ.get("LLAMA_CACHE", "tmp")
+
+    cmd = [
+        str(exe),
+        "-hfr", params["hf_repo"],
+        "-hff", params["hf_file"],
+        "--ctx-size", "1024",
+        "--embd-output-format", "raw",
+        "--threads", "1",
+    ]
+
+    env = os.environ.copy()
+    env["LLAMA_CACHE"] = cache_dir
+    start = time.perf_counter()
+    result = subprocess.run(cmd, input=text, capture_output=True, text=True, env=env)
+    if result.returncode != 0:
+        print(f"[worker {i}] stderr:\n{result.stderr}")
+        raise AssertionError(f"embedding run {i} failed (code {result.returncode})")
+    return time.perf_counter() - start
+
+
+@benchmark(n=1)
+def test_embedding_concurrent_invocations():
+    params = ensure_model_downloaded()
+    text = " ".join(["concurrency"] * 128)
+    n_workers = 4
+    with ProcessPoolExecutor(max_workers=n_workers) as pool:
+        futures = [pool.submit(run_one, (i, params, text)) for i in range(n_workers)]
+        times = [f.result() for f in as_completed(futures)]
+    avg = statistics.mean(times)
+    print(f"[concurrency] {n_workers} parallel runs: mean={avg*1000:.1f} ms")
+
+
+@benchmark(n=1)
+def test_embedding_large_model_logging_stress():
+    """Optional stress test using larger model for stdout/mutex path."""
+    large_model = {
+        "hf_repo": "TheBloke/Mistral-7B-Instruct-v0.2-GGUF",
+        "hf_file": "mistral-7b-instruct-v0.2.Q4_K_M.gguf",
+    }
+    text = " ".join(["benchmark"] * 8192)
+    out = run_embedding(text, "raw", params=large_model)
+    floats = [float(x) for x in out.split()]
+    assert len(floats) >= 1024
+
+
+def test_embedding_invalid_flag():
+    """
+    Invalid flag should produce a non-zero exit and a helpful error message.
+    Ensures CLI argument parsing fails gracefully instead of crashing.
+    """
+    repo_root = Path(__file__).resolve().parents[2]
+    exe = repo_root / "build/bin/llama-embedding"
+    assert exe.exists(), f"Missing binary: {exe}"
+
+    # Pass an obviously invalid flag to trigger error handling.
+    result = subprocess.run(
+        [str(exe), "--no-such-flag"],
+        capture_output=True,
+        text=True,
+    )
+
+    # Must return non-zero and print something meaningful to stderr.
+    assert result.returncode != 0, "Expected non-zero exit on invalid flag"
+    stderr_lower = result.stderr.lower()
+    assert (
+        "error" in stderr_lower
+        or "invalid" in stderr_lower
+        or "unknown" in stderr_lower
+    ), f"Unexpected stderr output: {result.stderr}"