exp-b7 with a test-time-training (TTT) memory in every layer (2 heads x 512, channeltanh gate). Only the TTT memory and the VAE/action input-output projections are trained. Recipe: wmttt/train/recipeformal2pairsb7ttthistory98gpuq4ttt10x.toml. MAST run formal2pairsb7ttthistory98gpuq4ttt10xdroid2cam (SyncWorld-dev wmttt, 2026-10-10/11). is fresh and LR warms up over 100 updates, so the step counter restarts at 0. The backbone stays frozen. the demonstration and the other the query. Both third-person cameras: each pair becomes two setups, one per camera, so every stream picks a camera 50/50, and its demonstration and query always use the same camera. Three components, mixed 1:1:1: - robocasa…
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.
The matched no-TTT control of the TTT run. It uses the same data, history-9 pair streams, losses, learning rates and number of updates, and trains the same VAE/action projections on the frozen backbone, without a TTT memory (tttenabled = false, pairchunkframes = 16). Recipe: wmttt/train/recipeformal2pairsb7nottthistory98gpuq4.toml. MAST run formal2pairsb7nottthistory98gpuq4droid2cam (SyncWorld-dev wmttt, 2026-10-10/11). is fresh and LR warms up over 100 updates, so the step counter restarts at 0. The backbone stays frozen. the demonstration and the other the query. Both third-person cameras: each pair becomes two setups, one per camera, so every stream picks a camera 50/50, and its…
We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively. Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially…
A text encoder for continuous diffusion language models (DLMs), distilled from the encoder of T5Gemma-2-270M. It gives the SoTA latent embeddings for continuous DLMs. From the paper Scaling and Distilling Text Embeddings for Better Diffusibility. A continuous diffusion language model generates text embeddings and decodes them to words. T5Gemma-2 provides a strong embedding space for this, but it keeps the words that could fill the same position far apart, and a generated embedding often lands between them and decodes problematically. The student is trained by minimizing the KL divergence between the probabilities that the same frozen T5Gemma-2 decoder predicts from the student's embeddings…
ComfyUI previews Krea 2 with Latent2RGB, a linear 16→3 color projection that looks blurry and blocky. This decoder is distilled from the Krea 2 VAE: it turns the sampler's latents into a real preview image in about 12 ms, with no need for the full VAE. 1. Download TAEKrea2AcademiaSD.safetensors to ComfyUI/models/vaeapprox/. 2. Install ComfyUI-KJNodes. 3. Add the Model Preview Override node between your Krea 2 model and the sampler. 4. In its tinyvae input, choose TAEKrea2AcademiaSD.safetensors. - Decoder only, there is no encoder. - Trained and tested with Krea 2 only. Tiny AutoEncoder (decoder tipo TAESD) para previsualizar Krea 2 en ComfyUI con nitidez y en tiempo real. Sustituye a…
is the previous three-component system plus a separately trained, small temporal CNN—not a new version of the spatial–temporal backbone. “Above” means the existing SAM-optimizer network + training-derived prototypes + original temporal CNN. The additional CNN uses the same labeled sample lists, fixed 32×32 crops and seed value 3407. We call it a control because we trained it alongside the attention-CNN and Transformer candidates to determine whether those architectural changes helped beyond simply training another model. For each spatial pixel it reads the sequence of ten Sentinel-2 spectral bands. Its convolutions operate along time, not across neighboring image pixels: It has 28,306…
Pretrained DiT decoders for TT-VidT (Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining, NeurIPS 2026). Paper · Each TT-VidT encoder pretraining run initialises its decoder from one of these; the training code downloads them automatically through DECODERPRETRAINED, e.g. "KBlueLeaf/TTVidT-decoders/pretrainvideoSqknorm". The decoders work in the 32x32x4 latent space of the frozen f8 frame VAE KBlueLeaf/latentmaid-vae and are conditioned on DINOv3 ViT-B/16 features (768-d). - ImageNet decoders reconstruct an ImageNet image from its DINOv3 class token. - Video decoders reconstruct a later video frame from its class token plus, through cross-attention, the spatial…
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.