Ziyang Ye 1, Junchao Huang 1,2, Evelyn Zhang 2, Zhihao Xie 1, Ruicheng Zhang 3, Boyao Han 1, Litao Ban 4, Ziye Wang 4, Xinting Hu 5, Shaoshuai Shi 4, Zhuotao Tian 2, Li Jiang 1,2† Project page | Code, so a strict loadstatedict works. The inference code needs the five weight files above and nothing else of this repository. EMA weights of the training stages, fp32, under the same names as the generator file; training from them starts from the exact weights. They also hold the visibility probe the training uses (8 tensors). The stage 2 checkpoint and the training code are added to this repository next. This fetches the inference files above, plus the Wan2.2 VAE, tokenizer and config.json from…
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.
12,283Models
3,510Datasets
290Papers
7,436Publishers
12,632Sourced relationships
Updated 2026-10-22 · How the library is built
12,283 models, sorted by most downloaded.
Base transformer weights and their matching camera adapter and LoRA for the WorldCrafter inference code. Place this directory beside WorldCrafter-Fast. Base reads the following shared folders from Fast: repencoder/, textencoder/, tokenizer/, vae/, and scheduler/. Base does not need Fast's transformer or adapter folders. The relative path is configured in inferenceconfig.json. Keep the Base-specific transformer/ and adapter/ together. SHA256SUMS covers the packaged Base files; shared component hashes are recorded in Fast's package.
Self-contained weights for WorldCrafter Fast image-to-video, text-to-video, and the interactive image-to-video demo. Use the WorldCrafter inference code and its pinned uv environment. The two transformer branches have independent DMD LoRA and camera adapters. The RepEncoder, text encoder, tokenizer, VAE, and scheduler are included here. No WorldCrafter-Base files are needed for Fast inference. Image-to-video uses five high-noise steps and one low-noise step per chunk. Text-to-video uses four high-noise and eight low-noise steps for the first chunk, then two high-noise and four low-noise steps. Compilation is optional with --enable-compile. inferenceconfig.json resolves shared components…
Wrench is a pruned, task-specific developer-tool SLM derived from Qwen3.6-35B-A3B. It contains 3,881,244,016 parameters and stays below the 4.25B parameter ceiling. This is a public experimental artifact. It is downloadable and reproducible, but it is not a claim that the final 4M retrieval-quality or MiniMax-parity gates have passed. The canonical distribution format is Hugging Face Safetensors. The package embeds the tokenizer hook, deterministic mechanical lookup runtime, long-context overlay, hash-bound package metadata, and the FreeToken launcher. It is intended to feel like one model directory, not a separately installed harness. The bundled launcher requires a compatible FreeToken…
ComfyUI-format conversions of Wuli-Art's MiniMax-H3 Turbo few-step audio+video distillations of - 4step/ — Wuli's 4-step adapter (turbo-4step/Wuli-MiniMax-H3-Turbo-4steps-V1.0-bf16.safetensors from Wuli-Art/MiniMax-H3-Turbo): a LoRA + 4 per-step output heads. - 8step/ — Wuli's 8-step transformer (turbo-8step/transformer from a full model with its 8 timesteps baked in, plus an int8 quant and an approximate LoRA. - 8step-ref2va/ — the same 8-step distill moved onto the Ref2VA model (reference-image conditioning): built as Ref2VA + (Wuli 8-step − FL2VA). A weight merge, not something Wuli trained. Untested in generation. - 8step-unlocked/, 8step-ref2va-unlocked/ — the same two 8-step int8…
This model is a fine-tuned version of Panga-Azazia/Wuro-Nucobiyo-Tanle. It has been trained using TRL. This model was trained with SFT.
This model is a fine-tuned version of Panga-Azazia/Wuro-Nucobiyo-Tanle. It has been trained using TRL. This model was trained with SFT.
This is an ONNX export of XPENG-AI/X-AuT, a Qwen3-ASR-0.6B model whose audio encoder was pruned from 18 to 14 layers. The files are laid out for onnxruntime-web on WebGPU. Demo: mrfakename/x-aut-webgpu. License: CC-BY-NC-4.0 (non-commercial), inherited from the original model. All credit for the model goes to XPENG-AI. This is an unofficial conversion. The decoder was built with the onnxruntime-genai builder (excludeembeds). 1. Log-mel features (Whisper-style, 128 bins, matching preprocessorconfig.json): - 16 kHz audio - nfft 400, hop 160, periodic Hann window, reflect pad - log10, clamped to max − 8, then (x + 4) / 4 - frames = len // 160 2. Encoder: run the encoder to get N audio…
This repository contains the X-Policy inference checkpoint used for the formal VLA-Arena leaderboard evaluation. X-Policy improves closed-loop temporal memory, task-routed action modeling, and replanning consistency. The repository contains params/, assets/, and CHECKPOINTMETADATA. Optimizer state and raw training state are intentionally excluded. X-Policy was evaluated on all 33 VLA-Arena suite/level cells using a fixed replanning interval of five environment steps. - L0 / L1 / L2 cell-mean SR: 76.18% / 50.18% / 33.27% The task-level scores and episode counts are available in evaluation/results.json. This is an Orbax checkpoint for X-Policy. It is not a Transformers frompretrained…
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).
To achieve optimal performance, use the following settings: Thinking: temperature=1.0, topp=0.95, topk=20, minp=0.0, presencepenalty=0.0, repetitionpenalty=1.0 Instruct: temperature=0.7, topp=0.80, topk=20, minp=0.0, presencepenalty=1.5, repetitionpenalty=1.0 This model is a part of The XORTRON Criminal Computing project; an ongoing research experiment and exercise in AI safety and alignment. XORTRON models are intentionally developed to study, evaluate, and document the potential of advanced AI systems to facilitate real-world criminal activity, abuse, and other high-risk conduct. Because these models may produce unusually capable, operationally relevant, or otherwise sensitive outputs…
Multimodal LLM researcher — VLMs for engineering & medical domains, RL post-training, open benchmarks. 多模态大模型研究者：工业/医疗垂直领域 VLM、RL 后训练（DAPO/GRPO）、开源 Benchmark。 Open to Work — 欢迎研究合作 / 全职机会。 Benchmark + models for comprehensive mechanical engineering drawing understanding. A pioneering open-source medical language model covering the full training pipeline (base → SFT → RL). Llama-3.1-8B instruct models for finance, transport, aerospace, law, hospitality & more (BAAI/IndustryInstruction), plus IndustryVL-2B-Instruct for industrial multimodal tasks. Realistic evaluation dataset for scholarly paper retrieval systems — Dataset · arXiv:2507.15245 Multimodal LLMs (VLM) · Domain adaptation…
This repository contains the post-trained Xiaomi-Robotics-1 (XR-1) checkpoint for further post-training with the Xiaomi-Robotics-1 codebase. XR-1 is a Vision-Language-Action (VLA) foundation model for mobile manipulation, initialized from a model pre-trained on over 100K hours of embodiment-free UMI trajectories and subsequently post-trained on over 10K hours of cross-embodiment data. The reference environment requires Python >= 3.9, a CUDA GPU, transformers==4.57.1, and FlashAttention 2. The Qwen3-VL backbone is loaded from Qwen/Qwen3-VL-4B-Instruct. Clone the source repository and install the dependencies: Download modelstates.pt from the Files and versions tab and place it at: Download…
This repository contains the Hugging Face checkpoint used by Xiaomi-Robotics-1 for RoboCasa evaluation. It includes the model weights, custom Transformers model and processor code, tokenizer files, and RoboCasa action normalization statistics. The custom model and processor must be loaded with trustremotecode=True. For reproducible multi-GPU evaluation, use the standard eight-server launcher in the companion source repository. Use the companion Xiaomi-Robotics-1 source repository and follow evalrobocasa/README.md. The server launcher accepts either this Hugging Face repository ID after downloading it locally or a local snapshot path. Apache License 2.0. See LICENSE.
This repository contains the Hugging Face checkpoint used by Xiaomi-Robotics-1 for RoboCasa365 evaluation. It includes the model weights, custom Transformers model and processor code, tokenizer files, and RoboCasa365 action normalization statistics. The custom model and processor must be loaded with trustremotecode=True. For reproducible multi-GPU evaluation, use the standard eight-server launcher in the companion source repository. Use the companion Xiaomi-Robotics-1 source repository and follow evalrobocasa365/README.md. The server launcher accepts either this Hugging Face repository ID after downloading it locally or a local snapshot path. Apache License 2.0. See LICENSE.
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.