Status: training in progress. No weights are published yet — this card describes the recipe and the pilot results that motivate it. A ~1B masked-diffusion language model decoded with confidence-targeted steps, then spend a few extra passes rewriting only the tokens the model is least sure about. The point is inference cost. An autoregressive model needs one sequential forward pass per token. This one needs ~20 passes for a whole sequence, regardless of its length. Cost is K + R forward passes. One refill pass fixes any number of positions at once, because the model processes the whole sequence in parallel — that is what makes targeted repair cheaper than more denoising. Draft and refill are…
SAVRN Model Hub
Open-Weight Models
An open-weight model is an AI model whose trained weights are published for anyone to download. The weights are what the model learned in training. With a copy of them you can run the model on hardware you control and train it further on your own data.
Open weights are not the same as open source. Many publishers release the weights without the training data or code, and the license sets what you may do with the model. This library puts each model's full card, architecture, files, license and published evaluations on one page.
2,965Models
944Datasets
255Papers
1,872Publishers
5,247Sourced relationships
Updated 2026-09-19 · How the library is built
2,965 models, sorted by most downloaded.
Flash Vision-Language-Action Inference for Autonomous Driving FlashDrive accelerates Alpamayo 1.5 — one of NVIDIA's 10B-parameter vision-language-action models for autonomous driving — by 4.7× with no loss in accuracy, through streaming inference, DFlash speculative reasoning, ParoQuant W4A8 quantization, adaptive action caching, and torch.compile. This repository mirrors the weights of nvidia/Alpamayo-1.5-10B and is the base checkpoint of the FlashDrive stack. Loading it pulls the derived companions automatically: Install FlashDrive, then load this base checkpoint — the -PARO and -DFlash companions are fetched automatically: The first call per stream only prefills the KV cache and returns…
Speech emotion recognition for Russian over seven classes: anger, disgust, enthusiasm, fear, happiness, neutral, sadness. Fine-tuned from jonatasgrosman/wav2vec2-large-xlsr-53-russian on Aniemore/resd. Audio resampled to 16 kHz mono, clips capped at 12 s, normalized per utterance, padding masked. UA is macro-averaged recall, WA is accuracy, F1 is macro-averaged. All three test sets went through the same harness, so the rows are comparable to each other. The RESD split matches fold 1 of EmoBox bit for bit. The top entry there is WavLM-large at WA 56.47 / UA 55.87 / F1 55.82. These numbers are higher, but the training protocol differs — EmoBox freezes the encoder and trains a probe, this is a…
T5 is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised and supervised tasks and for which each task is converted into a text-to-text format. For more information, please take a look at the original paper. Paper: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer Authors: Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu You can use this model with Transformers pipeline.
Fine-tuned facebook/wav2vec2-base for audio classification of single drum/percussion sounds into 10 classes. - clap, conga, crash, cymbal, hat, kick, ride, rim, snare, tom - Trained on short, single-hit drum sounds. Performance may drop on long mixes, multiple overlapping sounds, or very different recording conditions.
CED are simple ViT-Transformer-based models for audio tagging, achieving sota performance on Audioset. Notable differences from other available models include: 1. Simplification for finetuning: Batchnormalization of Mel-Spectrograms. During finetuning one does not need to first compute mean/variance over the dataset, which is common for AST. 1. Support for variable length inputs. Most other models use a static time-frequency position embedding, which hinders the model's generalization to segments shorter than 10s. Many previous transformers simply pad their input to 10s in order to avoid the performance impact, which in turn slows down training/inference drastically. 1. Training/Inference…
The RT-DETRv2 model was proposed in RT-DETRv2: Improved Baseline with Bag-of-Freebies for Real-Time Detection Transformer by Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, Yi Liu. RT-DETRv2 refines RT-DETR by introducing selective multi-scale feature extraction, a discrete sampling operator for broader deployment compatibility, and improved training strategies like dynamic data augmentation and scale-adaptive hyperparameters. These changes enhance flexibility and practicality while maintaining real-time performance. This model was contributed by @jadechoghari with the help of @cyrilvallez and @qubvel-hf This is RT-DETRv2 consistently outperforms its predecessor across all…
https://huggingface.co/sshleifer/distilbart-cnn-6-6 with ONNX weights to be compatible with Transformers.js. If you haven't already, you can install the Transformers.js JavaScript library from NPM using: Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).
Dasheng (Deep Audio-Signal Holistic Embeddings), or “大声” ("great sound"), is a general-purpose audio encoder trained on a large-scale self-supervised learning task. Dasheng is designed to capture rich audio information across various domains, including speech, music, and environmental sounds. The model is trained on 272,356 hours of diverse audio data with 1.2 billion parameters, and exhibits significant performance gains on the HEAR benchmark. Dasheng outperforms previous works on CREMA-D, LibriCount, Speech Commands, VoxLingua, and competes well in music and environmental sound classification tasks. examplefinetuneesc50.ipynb demonstrates how to train a linear head on the ESC-50 dataset…
Model · Audio classification
ast-finetuned-audioset-14-14-0.443
Audio Spectrogram Transformer (AST) model fine-tuned on AudioSet. It was introduced in the paper AST: Audio Spectrogram Transformer by Gong et al. and first released in this repository. Disclaimer: The team releasing Audio Spectrogram Transformer did not write a model card for this model so this model card has been written by the Hugging Face team. The Audio Spectrogram Transformer is equivalent to ViT, but applied on audio. Audio is first turned into an image (as a spectrogram), after which a Vision Transformer is applied. The model gets state-of-the-art results on several audio classification benchmarks. You can use the raw model for classifying audio into one of the AudioSet classes. See…
YOLOv8 is the eighth version of the You Only Look Once (YOLO) object detection algorithm. It excels in speed and accuracy, making it an ideal choice for real-time applications. The YOLOv8 model provided here has been fine-tuned on a diverse dataset of handwritten texts to improve its specificity in detecting handwritten content as opposed to typed or printed materials. The final IoU=0.98 The IoU during training
RF-DETR is a real-time detection transformer family introduced in RF-DETR: Neural Architecture Search for Real-Time Detection Transformers by Robinson et al. and integrated in Transformers via PR #36895. RF-DETR is an end-to-end object detection model that combines ideas from LW-DETR and Deformable DETR: a DINOv2-with-registers style ViT backbone (with an RF-DETR windowing pattern for efficient attention), a multi-scale projector between encoder and decoder, and a multi-scale deformable DETR decoder for fast convergence and strong accuracy–latency tradeoffs. You can use the raw model for object detection. See the model hub to look for all available RF-DETR models. Here is how to use this…
This model is a fine-tuned version of facebook/wav2vec2-base-960h for Speech Emotion Recognition (SER). It has been trained using a Frozen Feature Extractor strategy to preserve the model's acoustic understanding while adapting to emotion detection. This approach ensures stable performance and prevents "Catastrophic Forgetting," achieving nearly 80% accuracy on the validation set. Update: The "Calm" and "Neutral" classes have been merged to improve classification consistency, resulting in 7 distinct emotion classes. The model was trained on a combined dataset of ~12,000 audio files from: The model classifies audio into one of the following emotions: 1. Angry 2. Disgust 3. Fear 4. Happy 5.…
MolmoAct2 is an open vision-language-action model for robot control. It builds on Molmo2-ER and attaches a flow-matching continuous action expert that conditions on the VLM key-value cache through a per-layer connection. This checkpoint is fine-tuned on the SO-100/101 mixture with absolute joint-pose control and annotated language instructions. It is intended for both further fine-tuning and SO-100/101 policy inference. Use this checkpoint for SO-100/101 inference or for further fine-tuning. Dataset normalization metadata is stored in normstats.json. pass normtag="so100so101molmoact2" at inference time. Continuous action prediction is the intended and recommended inference mode. Discrete…
The D-FINE model was proposed in D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement by Yansong Peng, Hebei Li, Peixi Wu, Yueyi Zhang, Xiaoyan Sun, Feng Wu This model was contributed by VladOS95-cyber with the help of @qubvel-hf This is the HF transformers implementation for D-FINE coco -> model trained on COCO obj365 -> model trained on Object365 obj2coco -> model trained on Object365 and then finetuned on COCO D-FINE, a powerful real-time object detector that achieves outstanding localization precision by redefining the bounding box regression task in DETR models. D-FINE comprises two key components: Fine-grained Distribution Refinement (FDR) and Global…
This model is a fine-tune checkpoint of DistilBERT-base-cased, fine-tuned using (a second step of) knowledge distillation on SQuAD v1.1. This model reaches a F1 score of 87.1 on the dev set (for comparison, BERT bert-base-cased version reaches a F1 score of 88.7).
This is a ported version of The base model is hubert-base-ls960, which is pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. For more information refer to SUPERB: Speech processing Universal PERformance Benchmark Emotion Recognition (ER) predicts an emotion class for each utterance. The most widely used ER dataset IEMOCAP is adopted, and we follow the conventional evaluation protocol: we drop the unbalanced emotion classes to leave the final four classes with a similar amount of data points and cross-validate on five folds of the standard splits. For the original model's training and evaluation instructions refer to the…
Model · Zero-shot classification
xtremedistil-l6-h256-zeroshot-v1.1-all-33
This model was fine-tuned using the same pipeline as described in the model card for MoritzLaurer/deberta-v3-large-zeroshot-v1.1-all-33 and in this paper. The foundation model is microsoft/xtremedistil-l6-h256-uncased. The model only has 22 million backbone parameters and 30 million vocabulary parameters. The backbone parameters are the main parameters active during inference, providing a significant speedup over larger models. The model is 25 MB small. This model was trained to provide a very small and highly efficient zeroshot option, especially for edge devices or in-browser use-cases with transformers.js. For usage instructions and other details refer to this model card…
Audio Spectrogram Transformer (AST) model fine-tuned on Speech Commands v2. It was introduced in the paper AST: Audio Spectrogram Transformer by Gong et al. and first released in this repository. Disclaimer: The team releasing Audio Spectrogram Transformer did not write a model card for this model so this model card has been written by the Hugging Face team. The Audio Spectrogram Transformer is equivalent to ViT, but applied on audio. Audio is first turned into an image (as a spectrogram), after which a Vision Transformer is applied. The model gets state-of-the-art results on several audio classification benchmarks. You can use the raw model for classifying audio into one of the Speech…
Cosmos-Embed1 is a joint video-text embedder tailored for physical AI. It can be used for text-to-video retrieval, inverse video search, semantic deduplication, zero-shot and k-nearest-neighbors (kNN) classification, and as a base model for video curation tasks. It has state-of-the-art (SOTA) performance on autonomous vehicle (AV) and robotics datasets, while maintaining competitive performance in general domains. A fine-tuned variant is also provided for video anomaly detection and classification. This model is ready for commercial use. The Cosmos-Embed1 release includes the following embedders: Note: while each checkpoint was optimized at a specific fixed resolution (and default to…
T5 is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised and supervised tasks and for which each task is converted into a text-to-text format. For more information, please take a look at the original paper. Paper: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer Authors: Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu You can use this model with Transformers pipeline.
Update Feb 14, 2025: Chronos-Bolt & original Chronos models are now available on Amazon SageMaker JumpStart! Check out the tutorial notebook to learn how to deploy Chronos endpoints for production use in a few lines of code. Update Nov 27, 2024: We have released Chronos-Bolt models that are more accurate (5% lower error), up to 250 times faster and 20 times more memory-efficient than the original Chronos models of the same size. Check out the new models here. Chronos is a family of pretrained time series forecasting models based on language model architectures. A time series is transformed into a sequence of tokens via scaling and quantization, and a language model is trained on these…
X-CLIP model (large-sized, patch resolution of 14) trained fully-supervised on Kinetics-400. It was introduced in the paper Expanding Language-Image Pretrained Models for General Video Recognition by Ni et al. and first released in this repository. This model was trained using 8 frames per video, at a resolution of 224x224. Disclaimer: The team releasing X-CLIP did not write a model card for this model so this model card has been written by the Hugging Face team. X-CLIP is a minimal extension of CLIP for general video-language understanding. The model is trained in a contrastive way on (video, text) pairs. This allows the model to be used for tasks like zero-shot, few-shot or fully…
Model Collections
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
Embedding models for retrieval
Sentence and document embedding models used to build retrieval systems. Dimension and sequence length matter more than size here, and both come from the publisher.
Collection · 4 entries
Models that fit on one accelerator
Models whose publisher-reported parameter count puts them within reach of a single accelerator at common precisions. Memory needed depends on precision and serving configuration, so treat the parameter count as the starting point, not the answer.
Collection · 6 entries
Open-weight text models worth knowing
Widely used open-weight language models, chosen because each one is a distinct family rather than a variant of the one above it. Selection, not a ranking.
Collection · 3 entries
Speech and audio models
Recognition and synthesis models, grouped so the two directions are easy to compare.
Open-Weight Models Explained
What is an open-weight model?
An AI model whose trained weights are published for anyone to download, so it can be run, tested and fine-tuned on hardware the user controls.
Is an open-weight model the same as open source?
Not always. Open weights means the trained model can be downloaded. Open source usually also means the training code and data are available and the license allows broad reuse. Many open-weight models release the weights only.
Can I use an open-weight model commercially?
It depends on the license. Apache 2.0 and MIT allow commercial use. Other licenses limit it, for example to non-commercial use or below a set number of users. Every model page here shows its license.
How much memory does an open-weight model need?
About two bytes per parameter at 16-bit precision, so a 7-billion-parameter model needs roughly 14 GB for its weights, plus memory for the context it processes. Each model page lists its parameter count and the size of its files.
Related SAVRN Research
The hub sits beside SAVRN's market data and infrastructure research: what models cost to run, and what it takes to run them.
SAVRN Index
What open models cost to run
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Data center trackers and maps
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
How the Model Hub is built
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.