diff --git a/docs/docs/configuration/object_detectors.md b/docs/docs/configuration/object_detectors.md
index 046c6363f1..eb7244d9ae 100644
--- a/docs/docs/configuration/object_detectors.md
+++ b/docs/docs/configuration/object_detectors.md
@@ -799,14 +799,10 @@ See the [installation docs](../frigate/installation.md#axelera-metis) for inform
 
 The integration runs inference through a single stream created with `create_inference_stream(network=...)`, which accepts a compiled model as a URL (downloaded into the model cache), a local file path, or an SDK model-zoo preset name. When no model is specified, the SDK model-zoo preset `yolox-s-coco-onnx` is used (no download); it can also be selected explicitly with `model.path: yolox-s-coco-onnx`.
 
+Models must be compiled for the Metis architecture — see [Building a model](/frigate/installation#building-a-model) for the full walkthrough.
+
 <ModelConfigDropdown detectorTitle="Axelera Metis" models={objectDetectorsModels.axelera.models} />
 
-:::info
-
-The Metis board needs models compiled for its architecture. Compile them from the Voyager model zoo with the SDK tooling, e.g. `axzoo build yolox_s_coco --backend axrelay --cal-source repr` (or compile a custom ONNX model with `axcompile`, see [Building a model](/frigate/installation#building-a-model)), producing a `.axm` file or a compiled model directory. Host the compiled model anywhere reachable, e.g. on Hugging Face, and point `model.path` at it. Plain strings in `model.path` are treated as SDK model-zoo preset names; anything else is treated as a file (downloaded from a URL or used in place).
-
-:::
-
 ### Downloading YOLO Models
 
 #### Axelera Metis
@@ -814,6 +810,4 @@ The Metis board needs models compiled for its architecture. Compile them from th
 The Metis board runs models compiled for its architecture (`.axm` files, or a compiled model directory). Two options are available:
 
 - **SDK model-zoo preset (no compilation or download needed):** the `yolox-s-coco-onnx` preset ships with the Voyager SDK (YOLOX, Megvii, Apache-2.0 licensed). Select it with `model.path: yolox-s-coco-onnx`.
-- **Compiled model:** build one with the SDK tooling, e.g. `axzoo build yolox_s_coco --backend axrelay --cal-source repr` (model zoo) or `axcompile -i model.onnx -o compiled/` (custom ONNX). See [Building a model](/frigate/installation#building-a-model).
-
-Host the compiled model somewhere Frigate can reach (a local path or an HTTP URL, e.g. on Hugging Face) and point `model.path` at it. With no `model.path`, the detector uses the SDK model-zoo preset `yolox-s-coco-onnx` by default (no download). Set `model.width` and `model.height` to the compiled model's input size.
+- **Compiled model:** see [Building a model](/frigate/installation#building-a-model) for how to compile one, then point `model.path` at it via a local path or an HTTP URL (e.g. on Hugging Face). Set `model.width` and `model.height` to the compiled model's input size.
diff --git a/frigate/detectors/plugins/axelera.py b/frigate/detectors/plugins/axelera.py
index ff7bbbe4a7..1d1153f9ed 100644
--- a/frigate/detectors/plugins/axelera.py
+++ b/frigate/detectors/plugins/axelera.py
@@ -24,12 +24,12 @@ Model configuration follows the hardware-backend convention:
   - `model.width` / `model.height`: input frame dimensions for the network.
 """
 
+import collections
 import logging
 import os
 import queue
 import threading
 import time
-import urllib.request
 from typing import Literal
 
 import numpy as np
@@ -38,6 +38,7 @@ from pydantic import ConfigDict
 from frigate.const import MODEL_CACHE_DIR
 from frigate.detectors.detection_api import DetectionApi
 from frigate.detectors.detector_config import BaseDetectorConfig
+from frigate.util.downloader import ModelDownloader
 
 logger = logging.getLogger(__name__)
 
@@ -48,10 +49,6 @@ DETECTOR_KEY = "axelera"
 # and cannot be referenced by Frigate (MIT), see blakeblackshear/frigate#10717.
 AX_DEFAULT_NETWORK = "yolox-s-coco-onnx"
 
-# Optional download URL for a compiled default model. When empty the SDK
-# model-zoo preset AX_DEFAULT_NETWORK is used instead (no download needed).
-AX_DEFAULT_URL = ""
-
 INFERENCE_TIMEOUT = 5.0
 MAX_ROWS = 20
 
@@ -60,37 +57,48 @@ MAX_ROWS = 20
 
 
 def is_url(value: str) -> bool:
-    return value.startswith(("http://", "https://", "www."))
+    """Return True for well-formed http(s) URLs (scheme required)."""
+    return value.startswith(("http://", "https://"))
+
+
+def _looks_like_path(value: str) -> bool:
+    """Heuristic: a value that should refer to a file on disk, not a preset."""
+    return "/" in value or "." in value
 
 
 def _resolve_network(model_path, model_url) -> str:
     """Resolve the value passed to ``create_inference_stream(network=...)``.
 
-    Returns a download URL -> cached local file / existing local file / SDK
-    model-zoo preset name, mirroring the resolution hailo8l performs for ``.hef``.
+    Returns a download URL cached as a local file / an existing local file /
+    an SDK model-zoo preset name, mirroring the resolution hailo8l performs
+    for ``.hef``.
     """
     if model_url and model_path:
         raise ValueError("`model.path` cannot be both a URL and a local path.")
 
-    if model_url or (not model_path and AX_DEFAULT_URL):
-        url = model_url or AX_DEFAULT_URL
-        name = os.path.basename(url) or AX_DEFAULT_NETWORK
+    if model_url:
+        name = os.path.basename(model_url) or AX_DEFAULT_NETWORK
         os.makedirs(MODEL_CACHE_DIR, exist_ok=True)
         destination = os.path.join(MODEL_CACHE_DIR, name)
         if not os.path.isfile(destination):
-            logger.info("axelera: downloading model from %s", url)
+            logger.info("axelera: downloading model from %s", model_url)
             try:
-                urllib.request.urlretrieve(url, destination)
+                ModelDownloader.download_from_url(model_url, destination)
             except Exception as exc:
                 raise RuntimeError(
-                    f"Failed to download model from {url}: {exc}"
+                    f"Failed to download model from {model_url}: {exc}"
                 ) from exc
         return destination
 
     if model_path:
-        if os.path.isfile(model_path):
+        if _looks_like_path(model_path):
+            if not os.path.isfile(model_path):
+                raise FileNotFoundError(
+                    f"Model file not found: {model_path}. If you meant to use "
+                    "an SDK model-zoo preset name, drop the file extension / path."
+                )
             return model_path
-        # Not an existing file: treat as an SDK model-zoo preset name.
+        # Bare name: SDK model-zoo preset.
         return model_path
 
     return AX_DEFAULT_NETWORK
@@ -120,13 +128,12 @@ class _AxeleraInference:
     never starves and is not torn down by its internal inference timeout during
     quiet periods (e.g. a camera with no motion).
 
-    Results are NOT strictly paired to request ids: per-frame pairing
-    desynchronises in production once keep-alive results interleave with real
-    frames (a request can receive a zero-frame result, or a real result gets
-    attributed to the wrong frame). Instead the pump keeps the most recent
-    completed result, and ``run()`` pushes its frame then waits for a result
-    newer than the sequence observed at call time and returns it. Frigate only
-    needs the latest per-camera detections; it correlates objects across frames.
+    Results are STRICTLY paired to requests: every frame the feeder pushes into
+    the SDK registers its request id on a deque (in submission order), and each
+    stream result pops the next pending id and stores the outcome under it.
+    ``run()`` waits for its own id only, so a result can never be attributed to
+    a different camera or crop, and primer/keep-alive results (unregistered ids)
+    are dropped instead of being returned to a caller.
     """
 
     def __init__(
@@ -136,8 +143,9 @@ class _AxeleraInference:
         self._width, self._height = int(width), int(height)
         self._frame_q: queue.Queue = queue.Queue(maxsize=64)
         self._cond = threading.Condition()
-        self._latest_seq = 0
-        self._latest_result: np.ndarray | None = None
+        self._pending_ids: collections.deque = collections.deque()
+        self._results: dict[int, np.ndarray] = {}
+        self._next_id = 0
 
         if stream_factory is None:
             # lazy import: never import the SDK unless the detector is used
@@ -163,18 +171,26 @@ class _AxeleraInference:
 
     def _feed(self):
         # primer frames: the SDK init pre-buffers several frames before the
-        # source yields (validated on hardware).
+        # source yields (validated on hardware). Their ids are NOT registered,
+        # so their stream results are dropped by the pump.
         for _ in range(3):
             yield np.zeros((self._height, self._width, 3), np.uint8)
         while True:
             try:
-                frame = self._frame_q.get(timeout=1.0)
+                frame_id, frame = self._frame_q.get(timeout=1.0)
             except queue.Empty:
                 # keep-alive: no request pending, keep the SDK stream fed so the
-                # inference timeout never kills it during quiet periods
-                frame = np.zeros((self._height, self._width, 3), np.uint8)
+                # inference timeout never kills it during quiet periods. The
+                # None id means this frame's result is dropped by the pump.
+                frame_id, frame = None, np.zeros(
+                    (self._height, self._width, 3), np.uint8
+                )
             if frame is None:  # shutdown sentinel from stop()
                 return
+            if frame_id is not None:
+                # submission order == result order: the pump pops this id when
+                # the SDK yields the result for this frame.
+                self._pending_ids.append(frame_id)
             yield frame
 
     def _detection_rows(self, frame_result) -> list:
@@ -184,7 +200,15 @@ class _AxeleraInference:
         them by the region size, pixel boxes are discarded as outside the frame
         (zero events despite working detections).
         """
-        dets = getattr(frame_result, "detections", None) or []
+        dets = getattr(frame_result, "detections", None)
+        if dets is None:
+            return []
+        try:
+            # some SDK builds return a numpy array; `or []` on an array raises
+            # ValueError (ambiguous truth value) with 2+ detections
+            dets = list(dets)
+        except TypeError:
+            dets = [dets]
         rows = []
         for det in dets:
             label = getattr(getattr(det, "label", None), "name", None)
@@ -193,7 +217,6 @@ class _AxeleraInference:
             class_id = self._labelmap.get(str(label).strip())
             if class_id is None:
                 continue
-            score = 0.0
             raw_box = getattr(det, "box", None)
             if raw_box is None:
                 continue
@@ -211,8 +234,8 @@ class _AxeleraInference:
                 (
                     class_id,
                     score,
-                    y0 / self._height,
-                    x0 / self._width,
+                    max(0.0, y0) / self._height,
+                    max(0.0, x0) / self._width,
                     min(y1, self._height) / self._height,
                     min(x1, self._width) / self._width,
                 )
@@ -225,16 +248,20 @@ class _AxeleraInference:
         try:
             for frame_result in self._stream:
                 out = _pad_rows(self._detection_rows(frame_result))
-                with self._cond:
-                    self._latest_seq += 1
-                    self._latest_result = out
-                    self._cond.notify_all()
+                if self._pending_ids:
+                    # FIFO: the stream returns results in submission order, so
+                    # the oldest pending id belongs to this result.
+                    frame_id = self._pending_ids.popleft()
+                    with self._cond:
+                        self._results[frame_id] = out
+                        self._cond.notify_all()
         except Exception:
             logger.exception(
                 "axelera: inference stream failed; detector requires restart"
             )
 
     def run(self, tensor_input, timeout: float = INFERENCE_TIMEOUT) -> np.ndarray:
+        """Run one inference and return this frame's detections (see detect_raw)."""
         if tensor_input is None or getattr(tensor_input, "ndim", 0) < 3:
             raise ValueError(
                 f"Unexpected frame shape: {getattr(tensor_input, 'shape', tensor_input)}"
@@ -246,36 +273,45 @@ class _AxeleraInference:
             if frame.shape[0] != 1:
                 raise ValueError(f"Unexpected batch dimension in frame: {frame.shape}")
             frame = frame[0]
-        frame = np.ascontiguousarray(frame)
+        # Real copy: the shared-memory camera buffer is unmapped by Frigate as
+        # soon as detect_raw returns, and our queue can hold the frame past that
+        # point (timeout path, backed-up queue). NP views would let the SDK read
+        # a torn/unmapped segment.
+        frame = np.array(frame, dtype=np.uint8, order="C", copy=True)
+
         if not self._thread.is_alive():
             # The pump is gone (stream wedged): never block on the queue put,
             # the SDK side is not draining it and the queue would fill up,
             # permanently wedging this detector.
             raise RuntimeError("Axelera inference thread stopped; restart required.")
+
+        # Every submitted frame gets a unique id; the pump stores the result
+        # under this id, so we can wait for OUR frame's outcome exclusively.
+        self._next_id += 1
+        frame_id = self._next_id
         try:
-            self._frame_q.put_nowait(frame)
+            self._frame_q.put((frame_id, frame), timeout=2.0)
         except queue.Full:
-            # The SDK side is not draining: drop the frame instead of blocking
-            # forever, then fall through to the latest-result wait below.
-            logger.debug("axelera: dropping frame, inference queue is full")
+            raise RuntimeError(
+                "Axelera inference queue is full; restart required."
+            ) from None
 
         deadline = time.monotonic() + timeout
         with self._cond:
-            start_seq = self._latest_seq
-            while self._latest_seq <= start_seq:
+            while frame_id not in self._results:
                 remaining = deadline - time.monotonic()
                 if remaining <= 0:
                     break
                 self._cond.wait(remaining)
-            if self._latest_seq > start_seq:
-                result = self._latest_result
-                if result is not None:
-                    return result
-        if not self._thread.is_alive():
+            if frame_id in self._results:
+                result = self._results.pop(frame_id)
+                return result
+        if self._thread.is_alive():
             raise RuntimeError(
-                "Axelera inference thread stopped; restart required."
-            ) from None
-        return np.zeros((MAX_ROWS, 6), np.float32)
+                "Axelera inference stream returned no result for the submitted "
+                f"frame within {timeout:.1f}s (stream stall); restart required."
+            )
+        raise RuntimeError("Axelera inference thread stopped; restart required.")
 
     def stop(self):
         # cleanup must not raise
@@ -298,14 +334,13 @@ class AxeleraDetector(DetectionApi):
     type_key = DETECTOR_KEY
 
     def __init__(self, detector_config: "AxeleraDetectorConfig"):
-        super().__init__(detector_config)
-
         model = detector_config.model
         if model is None:
             raise ValueError(
                 f"type={DETECTOR_KEY}: the `model:` block is required "
                 "(path to a compiled .axm, URL, or SDK model-zoo preset name)."
             )
+        super().__init__(detector_config)
 
         self.width = model.width
         self.height = model.height
@@ -337,9 +372,16 @@ class AxeleraDetector(DetectionApi):
         )
 
     def detect_raw(self, tensor_input) -> np.ndarray:
+        """Run the detector on one RGB frame and return (20, 6) float32 rows.
+
+        Each row is (class_id, score, ymin, xmin, ymax, xmax) in NORMALIZED
+        coordinates [0..1] relative to the model input size. Unused rows are
+        zero-filled. Detections are strictly paired to THIS frame's inference.
+        """
         return self._inference.run(tensor_input)
 
     def close(self):
+        """Stop the inference stream and release the Metis device."""
         self._inference.stop()
 
     def __del__(self):
diff --git a/frigate/stats/util.py b/frigate/stats/util.py
index af3a0b7892..f89118557a 100644
--- a/frigate/stats/util.py
+++ b/frigate/stats/util.py
@@ -344,7 +344,9 @@ async def set_npu_usages(config: FrigateConfig, all_stats: dict[str, Any]) -> No
             # Axelera Metis: temperature from the board controller.
             # There is no utilization readout without the axsystemserver
             # broker, which cannot run alongside a live detector.
-            metis_temp = get_axelera_board_temp()
+            # axcmd spawns can take seconds on a busy board; this stats task
+            # runs in the async event loop, so run the blocking read in a thread.
+            metis_temp = await asyncio.to_thread(get_axelera_board_temp)
 
             if metis_temp is not None:
                 stats["axelera"] = {"mem": "-%", "temp": metis_temp}
diff --git a/frigate/util/services.py b/frigate/util/services.py
index 8542eb6f91..20c610dcf9 100644
--- a/frigate/util/services.py
+++ b/frigate/util/services.py
@@ -800,12 +800,27 @@ def get_axcl_npu_stats() -> dict[str, str | float] | None:
     return None
 
 
+_board_temp_cache: tuple[float, float] | None = None
+_BOARD_TEMP_TTL = 5.0
+
+
 def get_axelera_board_temp() -> float | None:
     """Get the Axelera Metis board temperature through the SDK's axcmd tool.
 
     The read goes through the board controller and does not open an inference
-    stream, so it is safe to call while a detector is running.
+    stream, so it is safe to call while a detector is running. Results are
+    cached for a few seconds so the stats cycle (detector temperature + NPU
+    usage run in the same tick) spawns axcmd at most once.
+
+    NOTE: this function is blocking (up to ~5s on a hung board). Callers in
+    async contexts must wrap it in ``asyncio.to_thread``.
     """
+    global _board_temp_cache
+
+    now = time.monotonic()
+    if _board_temp_cache is not None and now - _board_temp_cache[0] < _BOARD_TEMP_TTL:
+        return _board_temp_cache[1]
+
     axcmd_path = None
 
     for candidate in (
@@ -826,6 +841,7 @@ def get_axelera_board_temp() -> float | None:
             capture_output=True,
             text=True,
             timeout=5,
+            check=False,
         )
     except (OSError, sp.TimeoutExpired):
         return None
@@ -833,16 +849,21 @@ def get_axelera_board_temp() -> float | None:
     if result.returncode != 0:
         return None
 
-    match = re.search(r"=\s*([\d.]+)", result.stdout)
+    # anchor on the board-controller line ("N: temp_sensorX@YY = <value>") so
+    # unrelated "= <number>" output (firmware/version preambles) is not picked up
+    match = re.search(r"temp_sensor[^=\n]*=\s*([\d.]+)", result.stdout)
 
     if match is None:
         return None
 
     try:
-        return round(float(match.group(1)), 1)
+        temp = round(float(match.group(1)), 1)
     except ValueError:
         return None
 
+    _board_temp_cache = (now, temp)
+    return temp
+
 
 def try_get_info(f, h, default="N/A", sensor=None):
     try: