English | 中文 A 143M-parameter bidirectional Chinese↔English translation model, instruction-tuned on the Haidass1.5-143M base — the strongest zh⇄en translator at this scale among general chat-architecture models. Drafter-143M: a control model with identical configuration, data and training recipe, except that it starts from random initialization instead of the pretrained base — used to quantify the contribution of base-model pretraining. OPUS-MT models are single-directional — one independent 78M model per direction; "-" marks directions a model does not serve. The same models re-evaluated on FLORES+ devtest (released 2026; zero overlap with dev): devtest sentences do not overlap with dev.…
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.
2,760Models
859Datasets
254Papers
1,692Publishers
5,040Sourced relationships
Updated 2026-09-18 · How the library is built
2,760 models, sorted by most downloaded.
Expert-paged build of Vontra/Qwen3.8-Flash-Next-MLX-4bit. The weights that are read a fraction at a time live in their own containers, so a machine loads what it needs rather than all Total 105.46 GiB. Of that, 103.94 GiB is the source build, whose bytes moved into containers rather than being copied, and 1.52 GiB is the draft head, which no published build of this model carries. Where the weights fit they are filled from experts.bin and the model runs the stock path at stock speed; where they do not, they stream from disk. Reading the machine decides that, not a flag. To override that: GBXPAGING=off holds the experts resident, GBXPLE=off holds the n-gram table resident. Checked at build…
Q4KM GGUF converted from dealignai/GLM-5.3-UNCENSORED-FP8. The source is a 753B-parameter mixture-of-experts model. See File folder for different Quants See the https://huggingface.co/zai-org/GLM-5.3 for usage. Thanks Z.ai for open sourcing this great model. Credit for the modified source weights belongs to dealignai. This repository ONLY provides the GGUF conversion.
VideoMAE model pre-trained on Something-Something-v2 for 2400 epochs in a self-supervised way. It was introduced in the paper VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training by Tong et al. and first released in this repository. Disclaimer: The team releasing VideoMAE did not write a model card for this model so this model card has been written by the Hugging Face team. VideoMAE is an extension of Masked Autoencoders (MAE) to video. The architecture of the model is very similar to that of a standard Vision Transformer (ViT), with a decoder on top for predicting pixel values for masked patches. Videos are presented to the model as a sequence of…
This model is a fine-tuned version of MCG-NJU/videomae-base on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - trainingsteps: 370 - Transformers 5.16.1 - Pytorch 2.14.0+cu126 - Datasets 5.0.1 - Tokenizers 0.23.2
COLLISION-1B is the official primary flagship model of the COLLISION ecosystem. Packing 999,376,128 parameters (~1.00B) into an optimized 24-layer transformer architecture, it delivers rich contextual reasoning, full 1,024-token context capacity, and state-of-the-art hybrid NLP capabilities with grounded web and local retrieval. Run everything in your browser on free Google Colab in under 10 seconds: Clone this repository and run pure PyTorch inference directly: COLLISION features a complete, zero-latency NLP pipeline: COLLISION includes a full dual-process cognitive architecture featuring non-linear Graph-of-Thoughts (GoT) and Hegelian Dialectics: The flagship features an advanced…
A neural autoposer for full-body character rigs. Give it a handful of handles — where a hand should be, where the feet are planted, which way the head looks — and it returns a complete, natural pose for a 30-joint skeleton. It is what makes posing in the Animatica Blender add-on feel like moving a character rather than rotating thirty bones. It is a poser, not an animator: one pose per call, a few milliseconds each, no motion model and no temporal state. Ask it once per frame and you have an editable animation; ask it once and you have a pose. Any number of effectors, in any combination. Each carries a tolerance — metres of slack — so a handle can be an exact pin or a loose suggestion the…
For usage instructions follow openai/whisper-large-v3. Large-v3 finetune trained as a baseline with smaller checkpoints in progress. Expecting worse long form and equal short form. Benchmarks. Has occasional repetition issue compared to previous models but achieves competitive/SOTA CER across all tested sets.
This repository contains the 6B model of the paper InternVideo2 in stage 2. Code: https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/multimodality Please refer to https://github.com/OpenGVLab/InternVideo/blob/main/InternVideo2/multimodality/INSTALL.md
A masked diffusion language model adapted from Qwen3-1.7B. The backbone is full attention, and every layer is made bidirectional. It is the control model in the paper's matched comparison against the hybrid dQwen3.5-2B. This is a base model, with no instruction tuning. Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models. Code: https://github.com/AntonXue/dQwen Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1). generate decodes the whole canvas at once, committing positions above a confidence threshold (tau=0.9); pass blocklength=32 for left-to-right block decoding, or tau=None, stepsperblock=k for a fixed budget. The 50B-token…
This model is a fine-tuned version of MCG-NJU/videomae-base on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - trainingsteps: 370 - Transformers 5.16.1 - Pytorch 2.11.0+cu128 - Datasets 5.0.1 - Tokenizers 0.23.1
Model · Video classification
videomae-huge-finetuned-kinetics
VideoMAE model pre-trained for 1600 epochs in a self-supervised way and fine-tuned in a supervised way on Kinetics-400. It was introduced in the paper VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training by Tong et al. and first released in this repository. Disclaimer: The team releasing VideoMAE did not write a model card for this model so this model card has been written by the Hugging Face team. VideoMAE is an extension of Masked Autoencoders (MAE) to video. The architecture of the model is very similar to that of a standard Vision Transformer (ViT), with a decoder on top for predicting pixel values for masked patches. Videos are presented to…
A masked diffusion language model adapted from Qwen3.5-2B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning. Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models. Code: https://github.com/AntonXue/dQwen Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1). generate decodes the whole canvas at once, committing positions above a confidence threshold (tau=0.9); pass blocklength=32 for left-to-right block decoding, or tau=None, stepsperblock=k for a fixed budget. The 50B-token checkpoint from the paper is…
A masked diffusion language model adapted from Qwen3.5-0.8B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning. Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models. Code: https://github.com/AntonXue/dQwen Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1). generate decodes the whole canvas at once, committing positions above a confidence threshold (tau=0.9); pass blocklength=32 for left-to-right block decoding, or tau=None, stepsperblock=k for a fixed budget. The 50B-token checkpoint from the paper is…
A frontier video understanding model developed by FAIR, Meta, which extends the pretraining objectives of VJEPA, resulting in state-of-the-art video understanding capabilities, leveraging data and model sizes at scale. The code is released in this repository. This is V-JEPA 2 ViT-g 384 model with video classification head pretrained on Diving 48 dataset. To run V-JEPA 2 model, ensure you have installed the latest transformers
This model is a fine-tuned version of MCG-NJU/videomae-large on the Deepfake Detection Challenge dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - lrschedulerwarmupratio: 0.1 - trainingsteps: 4470 - mixedprecisiontraining: Native AMP - Transformers 4.44.2 - Pytorch 2.5.0+cu121 - Datasets 3.1.0 - Tokenizers 0.19.1
Architecture: Qwen 3.6 35B-A3B (MoE) | Total Params: ~34.7B | Active Params: ~3B | Context: 262,144 native / 1,010,000 extensible | Base: Heretic (llmfan46) | Teacher: Claude Fable 5 | Type: Distilled + Abliterated MoE LLM A personal fork of llmfan46/Qwen3.6-35B-A3B-uncensored-heretic — an uncensored Heretic-style abliteration of Qwen/Qwen3.6-35B-A3B, the 35B-total / 3B-active mixture-of-experts multimodal base — repackaged as Janus-35B with Claude Fable 5 reasoning data in the teacher slot. Refusal-trained behavior is dialed back at the base layer. One-liner via Hugging Face (pulls a GGUF + this repo's root-level template / system / params files, including the tool-calling template — HF's…
This model is a fine-tuned version of microsoft/resnet-50 on the imagefolder dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 16 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 16 - lrschedulertype: cosine - numepochs: 20 - labelsmoothingfactor: 0.1 - Transformers 5.16.1 - Pytorch 2.11.0+cu128 - Datasets 4.8.5 - Tokenizers 0.23.1
A HuggingFace-format conversion of Meta AI's V-JEPA 2.1 ViT-g/16 video encoder and predictor, operating at 384x384 resolution. The weights are Meta's, copied without modification. This repository provides the transformers-compatible packaging plus a documented numerical validation against the original implementation. No prior HuggingFace port of this variant existed at the time of upload. The only structural change is that the fused QKV projection of each attention block is split into separate query / key / value matrices, following the convention used by transformers. This is a re-parameterization, not a change of weights. It is also convenient downstream: PEFT adapters apply to…
A masked diffusion language model adapted from Qwen3.5-4B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning. Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models. Code: https://github.com/AntonXue/dQwen Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1). generate decodes the whole canvas at once, committing positions above a confidence threshold (tau=0.9); pass blocklength=32 for left-to-right block decoding, or tau=None, stepsperblock=k for a fixed budget. The 50B-token checkpoint from the paper is…
Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded. Finetuned from turbo with pruned vocab, indices can be found in mapping.txt. Trained decoder only for 2^20 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps. Benchmarks. Competitive/SOTA on test sets, slightly better than 1.5B on short form, worse on long form. Also trained for…
This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
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.