NVIDIA Jetson Orin

A compact 450M-parameter VLA from Hugging Face, pretrained on community-collected LeRobot datasets and sized for fine-tuning and inference on consumer GPUs.

A 3B cross-embodiment robot foundation model from NVIDIA that combines a vision-language backbone with a diffusion-transformer action head and can be post-trained on a new robot with a modest amount of data.

A 93M-parameter transformer generalist policy with a diffusion action head, conditioned on language or goal images and intended to be fine-tuned to new robots, sensors, and action spaces.

An Action Chunking Transformer policy trained in LeRobot on human demonstrations of the simulated ALOHA bimanual cube-transfer task.

A Diffusion Policy checkpoint trained in LeRobot to push a T-shaped block into a target pose in the 2D PushT simulation.

Meta's self-supervised ViT-L video encoder, pretrained by predicting masked video regions in latent space rather than in pixels.

Compact variant of Meta's promptable segmentation model that returns object masks from points or boxes and can track them across video frames.

Open-vocabulary detector that draws boxes around objects named in a free-text prompt, without a fixed class list.

CLIP-based zero-shot detector from Google that localizes objects from text queries or from an example image crop.

Small real-time DETR-style detector for the 80 COCO classes that refines box edges as distributions for tighter localization.

Lightweight monocular model that predicts a relative depth map from a single RGB frame.

Plain vision-transformer model that estimates 2D human body keypoints for each person box produced by a separate detector.

Learned matcher that pairs SuperPoint keypoints between two images and stops early on easy pairs to save compute.

Smallest ViT from Meta's DINOv3 family, a self-supervised backbone whose dense patch features transfer to many vision tasks without fine-tuning.

Half-billion-parameter vision-language model from Hugging Face that answers questions about images and short video clips.

Compact, frequently updated vision-language model with built-in captioning, visual Q&A, object detection and pointing.

Smallest instruction-tuned model in Alibaba's Qwen3-VL line, covering image and video understanding, OCR and grounded object localization.

Google's 3B vision-language model trained on a mix of tasks, including captioning, visual Q&A, detection and segmentation, at 224px input.

Mobile-first Gemma variant that accepts text, image and audio input and runs with a memory footprint close to a 2B model by offloading per-layer embeddings.

A speech-input LLM that feeds Whisper encoder features through a learned projector into Llama 3.2 1B Instruct, so it can respond to spoken input without a separate transcription step. It replies in text, so pair it with a TTS voice.

A faster Whisper large-v3 derivative whose decoder is cut from 32 layers to 4, giving multilingual transcription in roughly 100 languages at a small accuracy cost (it was not fine-tuned for translation).

The small size of the Moonshine Streaming family, an English-only encoder-decoder recognizer that uses a sliding-window encoder so it can transcribe live audio with low latency on edge hardware.

A FastConformer encoder with a token-and-duration transducer decoder that transcribes 25 European languages with punctuation, capitalization and word-level timestamps.

A cache-aware FastConformer-RNNT streaming recognizer that covers 40 language-locales from one checkpoint via language-ID prompting, outputs punctuated text, and lets you pick chunk sizes from 80 ms up to about 1.1 s.

A decoder-only streaming English and French transcriber based on Kyutai's delayed-streams design, running about 0.5 s behind the audio and exposing a semantic voice-activity signal for end-of-turn detection.

GGML-format conversions of the OpenAI Whisper checkpoints, from tiny to large, for the whisper.cpp C/C++ inference engine.

An 82M-parameter text-to-speech model built on StyleTTS 2 components, shipped with a set of preset voices across several languages.

The official collection of ONNX voice models for the Piper offline TTS engine, covering many languages at several size and quality levels.

A 100M-parameter CPU-oriented TTS model from Kyutai with voice cloning from a reference clip and support for English plus five other European languages, designed to run faster than real time using two CPU cores.

A conversational speech model that produces Mimi audio codes from text and prior dialogue audio, so its prosody follows the flow of the conversation.

A family of 0.5B zero-shot voice-cloning TTS models whose repo holds the English model, which has an emotion-exaggeration control, alongside multilingual checkpoints covering 23 languages.

The smaller base checkpoint of Qwen3-TTS, a streaming text-to-speech model on a 12 Hz speech tokenizer that covers 10 languages and clones a voice from a roughly 3-second reference clip.

ONNX exports of the Silero voice activity detector in fp32, fp16 and 8-bit forms (the int8 file is about 0.6 MB) that mark which short audio frames contain speech.

An 8M-parameter multilingual turn-detection model, a Whisper Tiny encoder with a linear head, that predicts from raw audio whether the speaker has finished; the int8 CPU ONNX build is about 8.7 MB.

Pretrained openWakeWord models, small classifiers on top of a shared speech-embedding front end, that detect phrases such as 'hey jarvis' in a continuous audio stream.

The open speaker-diarization pipeline released with pyannote.audio 4.0, combining a segmentation model, speaker embeddings and PLDA-based clustering to label who spoke when, with better speaker counting than the 3.1 pipeline.

Fully open 3B language model with switchable reasoning mode, long context, tool calling and multilingual support, released with its training recipe.

Smallest dense model in the Qwen3 family, with optional thinking mode and function-calling support.

Tiny instruction-tuned Gemma 3 model meant to be fine-tuned for narrow jobs such as classification, extraction or command routing.