moondream 2.2.0__tar.gz → 2.4.0__tar.gz

{moondream-2.2.0 → moondream-2.4.0}/CHANGELOG.md

@@ -1,5 +1,21 @@
 # Changelog
 
+## 2.4.0
+
+- Added fast, memory-efficient local transcription with `moondream/parakeet-redux` on CPUs, Apple silicon, and CUDA, including pause-aware processing for long audio.
+
+## 2.3.0
+
+- Upgraded Photon local inference to `kestrel 0.7.2`.
+- Expanded fast Qwen 3.5 and Qwen 3.6 35B-A3B inference across NVIDIA
+  Ampere, Hopper, and Blackwell GPUs, including changing request concurrency.
+- Kept block-scaled Qwen prefill projections on the packed GPU path.
+- Corrected multimodal Qwen chat token accounting for conversations with
+  image inputs.
+
+See the underlying engine [0.7.2](https://github.com/m87-labs/kestrel/blob/main/CHANGELOG.md#072--2026-09-13)
+release notes for the complete engine changes.
+
 ## 2.2.0
 
 - Photon's optimized local inference now spans NVIDIA Ampere, Ada, Hopper, and
@@ -9,8 +25,6 @@
 - Added accelerated Whisper transcription on L4 and RTX 3090.
 - Reduced transcription latency and host overhead for Qwen ASR and Parakeet,
   including one-at-a-time workloads.
-- A standard `pip install moondream==2.2.0` now selects the matching optimized
-  Photon runtime and GPU kernels automatically.
 
 See the underlying engine [0.7.1](https://github.com/m87-labs/kestrel/blob/main/CHANGELOG.md#071--2026-09-09)
 release notes for the complete engine changes.

moondream-2.2.0/README.md → moondream-2.4.0/PKG-INFO

@@ -1,3 +1,20 @@
+Metadata-Version: 2.4
+Name: moondream
+Version: 2.4.0
+Summary: Official Python interface for Moondream Cloud and Photon local inference.
+Author: M87 Labs
+Author-email: contact@moondream.ai
+Requires-Python: >=3.10,<3.15
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Requires-Dist: kestrel (==0.8.0)
+Requires-Dist: pillow (>=10.4.0)
+Description-Content-Type: text/markdown
+
 # Moondream Python Client Library
 
 Official Python interface for [Moondream Cloud](https://moondream.ai/cloud) and Photon local inference on NVIDIA GPUs (Linux x86_64 / aarch64 or Windows) or Apple Silicon Macs.
@@ -30,7 +47,7 @@ Photon 2.1 includes these local model families:
 | Gemma 4 | E2B, E4B, 26B-A4B, and 31B base/instruction variants |
 | Whisper | Whisper large-v3-turbo transcription and English translation |
 | Qwen3-ASR | 0.6B and 1.7B transcription and forced alignment |
-| Parakeet TDT | 0.6B v3 transcription |
+| Parakeet TDT | 0.6B v3 transcription; [parakeet-redux](https://huggingface.co/moondream/parakeet-redux), its ternary version for CPUs, Apple silicon, and CUDA |
 
 Use `md.photon_models()` to inspect the exact registered identifiers in the installed
 release. The returned client reports `model_id`, `tasks`, and
@@ -106,8 +123,12 @@ model = md.vl(api_key="<your-api-key>", model="moondream3-preview/ft_id@step")
 qwen = md.photon("Qwen/Qwen3.5-4B")
 gemma = md.photon("google/gemma-4-E2B-it")
 speech = md.photon("openai/whisper-large-v3-turbo")
+speech = md.photon("moondream/parakeet-redux")               # CPU, Apple silicon or CUDA
 ```
 
+Photon picks CUDA when it is available, then Apple silicon, then the CPU; pass
+`device="cpu"`, `"mps"` or `"cuda"` to choose.
+
 Photon clients share matching local engines. Call `model.close()` when an
 application is finished with a client, or use
 `with md.photon("moondream3.1-9B-A2B") as model:`
@@ -324,6 +345,16 @@ Set `task="translate"` for English translation. Other options include
 `language`, `sample_rate`, `initial_prompt`, `condition_on_previous_text`,
 `clip_start_seconds`, `clip_end_seconds`, and model sampling `settings`.
 
+`moondream/parakeet-redux` is the ternary Parakeet: 178 MB of weights, 25
+languages, and local inference on CPUs, Apple silicon, and CUDA. It takes
+`timestamps` of `"none"`, `"segment"`, `"word"` or `"character"` and no
+language or prompt options.
+
+```python
+with md.photon("moondream/parakeet-redux") as speech:
+    print(speech.transcribe(audio="meeting.mp3")["text"])
+```
+
 ### Types
 
 | Type | Description |
@@ -340,3 +371,4 @@ Set `task="translate"` for English translation. Other options include
 - [Website](https://moondream.ai/)
 - [Playground](https://moondream.ai/playground)
 - [GitHub](https://github.com/vikhyat/moondream)
+

moondream-2.2.0/PKG-INFO → moondream-2.4.0/README.md

@@ -1,20 +1,3 @@
-Metadata-Version: 2.4
-Name: moondream
-Version: 2.2.0
-Summary: Official Python interface for Moondream Cloud and Photon local inference.
-Author: M87 Labs
-Author-email: contact@moondream.ai
-Requires-Python: >=3.10,<3.15
-Classifier: Programming Language :: Python :: 3
-Classifier: Programming Language :: Python :: 3.10
-Classifier: Programming Language :: Python :: 3.11
-Classifier: Programming Language :: Python :: 3.12
-Classifier: Programming Language :: Python :: 3.13
-Classifier: Programming Language :: Python :: 3.14
-Requires-Dist: kestrel (==0.7.1)
-Requires-Dist: pillow (>=10.4.0)
-Description-Content-Type: text/markdown
-
 # Moondream Python Client Library
 
 Official Python interface for [Moondream Cloud](https://moondream.ai/cloud) and Photon local inference on NVIDIA GPUs (Linux x86_64 / aarch64 or Windows) or Apple Silicon Macs.
@@ -47,7 +30,7 @@ Photon 2.1 includes these local model families:
 | Gemma 4 | E2B, E4B, 26B-A4B, and 31B base/instruction variants |
 | Whisper | Whisper large-v3-turbo transcription and English translation |
 | Qwen3-ASR | 0.6B and 1.7B transcription and forced alignment |
-| Parakeet TDT | 0.6B v3 transcription |
+| Parakeet TDT | 0.6B v3 transcription; [parakeet-redux](https://huggingface.co/moondream/parakeet-redux), its ternary version for CPUs, Apple silicon, and CUDA |
 
 Use `md.photon_models()` to inspect the exact registered identifiers in the installed
 release. The returned client reports `model_id`, `tasks`, and
@@ -123,8 +106,12 @@ model = md.vl(api_key="<your-api-key>", model="moondream3-preview/ft_id@step")
 qwen = md.photon("Qwen/Qwen3.5-4B")
 gemma = md.photon("google/gemma-4-E2B-it")
 speech = md.photon("openai/whisper-large-v3-turbo")
+speech = md.photon("moondream/parakeet-redux")               # CPU, Apple silicon or CUDA
 ```
 
+Photon picks CUDA when it is available, then Apple silicon, then the CPU; pass
+`device="cpu"`, `"mps"` or `"cuda"` to choose.
+
 Photon clients share matching local engines. Call `model.close()` when an
 application is finished with a client, or use
 `with md.photon("moondream3.1-9B-A2B") as model:`
@@ -341,6 +328,16 @@ Set `task="translate"` for English translation. Other options include
 `language`, `sample_rate`, `initial_prompt`, `condition_on_previous_text`,
 `clip_start_seconds`, `clip_end_seconds`, and model sampling `settings`.
 
+`moondream/parakeet-redux` is the ternary Parakeet: 178 MB of weights, 25
+languages, and local inference on CPUs, Apple silicon, and CUDA. It takes
+`timestamps` of `"none"`, `"segment"`, `"word"` or `"character"` and no
+language or prompt options.
+
+```python
+with md.photon("moondream/parakeet-redux") as speech:
+    print(speech.transcribe(audio="meeting.mp3")["text"])
+```
+
 ### Types
 
 | Type | Description |
@@ -357,4 +354,3 @@ Set `task="translate"` for English translation. Other options include
 - [Website](https://moondream.ai/)
 - [Playground](https://moondream.ai/playground)
 - [GitHub](https://github.com/vikhyat/moondream)
-

{moondream-2.2.0 → moondream-2.4.0}/moondream/photon_vl.py

@@ -5,6 +5,10 @@ import atexit
 import base64
 import json
 import os
+
+# Set before torch loads so idle OpenMP workers do not compete with Kestrel's CPU kernels.
+os.environ.setdefault("OMP_WAIT_POLICY", "passive")
+
 import queue
 import threading
 from concurrent.futures import TimeoutError as FutureTimeoutError
@@ -32,7 +36,12 @@ from .types import (
 
 
 def _default_photon_device() -> str:
-    """Choose the local Photon device when the caller does not specify one."""
+    """Choose the local Photon device when the caller does not specify one.
+
+    CUDA when it is available, Apple silicon's MPS when it is, and the CPU
+    otherwise. Models that do not run on the CPU report that themselves when
+    the engine starts; passing ``device=`` overrides the choice.
+    """
     if torch.cuda.is_available():
         return "cuda"
     if torch.backends.cuda.is_built():
@@ -42,10 +51,7 @@ def _default_photon_device() -> str:
         return "cuda"
     if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
         return "mps"
-    raise RuntimeError(
-        "Photon local inference needs a supported accelerator, but neither "
-        "CUDA nor Apple Silicon MPS is available in this Python environment."
-    )
+    return "cpu"
 
 
 def _image_to_bytes(image: Union[Image.Image, EncodedImage]) -> bytes:

{moondream-2.2.0 → moondream-2.4.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "moondream"
-version = "2.2.0"
+version = "2.4.0"
 description = "Official Python interface for Moondream Cloud and Photon local inference."
 authors = [ "M87 Labs <contact@moondream.ai>",]
 readme = "README.md"
@@ -21,4 +21,4 @@ reportMissingParameterType = false
 [tool.poetry.dependencies]
 python = ">=3.10,<3.15"
 pillow = ">=10.4.0"
-kestrel = "0.7.1"
+kestrel = "0.8.0"