Search public pages, research tools, and SAVRN solutions.
SAVRN Model Hub
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.
Updated 2026-10-18 · How the library is built
11,122 models, sorted by most downloaded.
Model · Text generation
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). - PEFT 0.20.0
AI Systems Engineer & Founder · Varanasi, India · experimentlab.in Building sovereign domain-specialized Small Language Models (SLMs) and edge-native reasoning systems. - 12 Live Colab Demos: Interactive Gradio web applications running on free Google Colab T4 GPUs. - 29 Open-Weight Models: Specialized models across Law, Math, Hardware, and STEM. - 5 Research Papers: Sovereign training recipes, MoE upcycling, DUS, and test-time reasoning on Kaggle. - 3 Curated Datasets: Over 1 billion tokens of clean STEM & Legal training data. Every demo launches an interactive Gradio chat interface connected to the real model running in 4-bit precision on a free Google Colab T4 GPU. No API keys or accounts…
A small Coca implementation for Classification, packaged with an explicit configuration and an initialization checkpoint. The small variant is a reproducible starting point, not a trained model release. - The Python file contains the model and runnable example or training entry point. - config.json records the generated architecture settings. - trainingargs.json records the default experiment recipe. - model.safetensors is a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint. - No benchmark score is claimed in this repository. The included configuration uses rmsprop with a step schedule. These are starting values in the script, not…
Sift is an AI-assisted SMS scam screening project designed to help international students review unfamiliar English messages. This repository contains its primary classification model, a trained scikit-learn pipeline that predicts Normal (0) or Scam (1) and outputs a score for the Scam class. The original labeled dataset has 500 examples: 250 Scam and 250 Normal. It includes SMS-style messages, adapted messages, and some source material from other communication channels; not every example was an independently observed real-world SMS. Normal: 238 synthetic or adapted examples based on official information from organizations, including ANZ, ITV, U.S. Bank, O2, and Unity Bank, plus 12…
Core ML conversion of Google's SigLIP 2 google/siglip2-base-patch16-256, with both towers: an image encoder and a text encoder, each returning an L2-normalized 768-d embedding. Labels are plain text chosen at run time, so it classifies images zero-shot. Google authored the Apache-2.0 source model; Fluid Inference converted it. The image encoder runs entirely on the Apple Neural Engine. fp16 weights, macOS 14 / iOS 17 or newer. config.json holds the preprocessing and scoring constants: trained without an attention mask, so the padding must match. an independent probability per label; softmax or argmax over cos picks one label. Embed the labels once and reuse them for every image. ImageNet-1k…
Selected Group 1 epoch-200 model weights (manifest.json), plus historical flagship and scaling-law weights (manifestflagshipscaling.json). For the large-model batch, epoch = step / 1251. Historical 40x1024 is epoch440; 28x1152 is epoch80. L-wide/deep are epoch200; H no-dropout includes epoch200 and epoch280; H projdrop0.2 is epoch80. Latest is not a claim of best FID. The H drop0.1 attempt has no checkpoint and is not included. Public derivatives retain raw model, EMA(s), and allowlisted architecture configuration only. No optimizer, RNG, credentials, W&B identity or local paths are included. These are inference weights, not full training-resume backups. Original checkpoints remain…
Glade is a native Swift library for on-device speech inference on Apple silicon. This repository supplies converted model assets and configuration for Glade's model-specific host runtime. Glade's runtime source is currently private. The official Silero 16 kHz VAD converted for Glade using Apple Core AI APIs on macOS 27 and iOS 27. The bundle batches eight consecutive 32 ms frames (256 ms), while preserving the original 128-dimensional LSTM state and 64-sample audio context across calls. Use CPU-only specialization for this small recurrent model. Each input audio tensor has shape [8, 1, 1, 576], containing the preceding 64 samples and current 512 samples for each frame. h and c are [1, 128…
Publicly shared presets, lorebooks, characters and whatnot. Tested extensively with Unsloth's Gemma4-31B-IT QAT Q4KXL, MTP enabled, 32K BF16 context. Model download: link Mistral Small 3.2 DeepSeek (v3.2, v4, v4.1) GLM (4.5/4.6/4.7/5.0/5.2/5.3)...and many others Builds off Voyage-v4-exp2. -...and many more things Tested extensively with Unsloth's Gemma4-31B-IT QAT Q4KXL, MTP enabled, 32K BF16 context. Model download: link Mistral Small 3.2 DeepSeek (v3.2, v4, v4.1) GLM (4.5/4.6/4.7/5.0/5.2/5.3)...and many others Experimental releases. Tested with Unsloth's Gemma4-31B-IT QAT Q4KXL, MTP enabled, 32K BF16 context. Model download: link Might work with other non-Gemma4 models. A playground of…
Experimental PEFT adapter archive. The amber-cedar snapshot uses Qwen/Qwen3-30B-A3B-Instruct-2507, revision 0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe, with PEFT 0.20.0.
Unsupervised SimCSE (Gao, Yao & Chen, EMNLP 2021, Eq. 1) on top of bert-base-uncased, trained on sentences from a 100k-record SNLI subset as a course assignment (U2T02, Trends in Data Science). It maps a sentence to a 768-dimensional, L2-normalized vector; compare vectors with cosine similarity. Team: Julio Cesar De Aquino Castellanos, Gustavo Alexander Fuentes Marin, Valeria Nicol Hernandez Leon, Ricardo Daniel Horta Sanchez, Jose Angel Pech Xool, Lorena Danae Perez Lopez. This is a reproduction under a reduced budget, not the paper's released checkpoint: it is 8.5 Spearman points below the paper on STS-B test (see Limitations). The supervised companion model is…
Unsupervised SimCSE sentence encoder trained from bert-base-uncased on a 100k-record subset of SNLI, as a course replication of Gao, Yao & Chen (EMNLP 2021). It maps an English sentence to a 768-dimensional vector; compare sentences with cosine similarity. Team (Universidad Politécnica de Yucatán, U2T02): Acosta Castellanos Bianca Alexandra, Canche Chuc Angel Rivaldo, Ku Aguilar Russel, Sanchez Novelo Damian, Velasco Martin Jonathan Abisai. 165,528 unique sentences (premises and hypotheses, whitespace-stripped) from snlitrain100k.jsonl, a 100k-record subset of the SNLI train split. Labels are not used: each sentence is its own positive, the two views differ only by dropout. The paper…
Public source distribution for SimCT / KDFlow experiments. This repository stores a Git bundle, not a direct Git branch mirror. The HF Git server rejects ordinary historical image blobs. The bundle retains the original code commit and complete Git history. - Runtime, checkpoints, datasets and untracked remoteartifacts are excluded. Download simct-b200-portable.bundle and source-manifest.json from the same pinned HF revision. Verify SHA-256 against the manifest before Git operations. First checkout in a new directory: For a subsequent complete bundle at the same absolute path, with simct-code checked out on the same branch and no conflicting changes: A Git bundle is a read-only transport…
Design × Engineering × AI. I build real, on-device systems from real data — space-tech tools, private AI, and honest engineering — and put the work where anyone can use it. - The Universe / Satellite Engine — a time-accurate, real-data solar system + live satellite orbits, built from NASA/JPL/ESA/CelesTrak data. Truth over spectacle. → sinhaankur.com/lab/celestial - Vera — a private, on-device AI twin of someone you love: local LLM, private memory, on-device voice. Nothing leaves your machine. Part of the Unhosted family. - Unhosted — pooling the hardware you own into a private inference cluster. AI that lives where you do. - rag-engine — a small, fully on-device Retrieval-Augmented…
Public reproducibility artifacts for the seed-42 SIP program: Matched-IP25, Augmented CleanReplay, and the controlled codebook experiment. Work is ongoing; files being present do not establish a scientific success or a completed program. Adapters require the exact base model/revision specified in their recipe and adapter configuration. Base weights are not included. Follow the corresponding base model's license and use restrictions; this repository does not relicense third-party materials. Model outputs are research data and may contain unwanted, harmful, invalid, or otherwise unreliable advice. Do not use these research checkpoints as production assistants. - preserved/worker1 and…
This model was converted to GGUF format using Unsloth. - gemma-4-E2B-it.BF16-mmproj.gguf - gemma-4-E2B-it.Q4KM.gguf
This model classifies facial skin images into 6 common dermatological conditions using a fine-tuned EfficientNetV2B0 architecture. - Evaluated on a balanced, augmented custom dataset of real-world dermatological images. - Confusion Matrix indicates strong separation for all classes, with minor overlap in visually similar conditions. - Global Average Pooling + Dense(512, ReLU) + Dropout(0.4) + Dense(6, Softmax) - Fine-tuned with learning rate scheduling and layer unfreezing Softmax probabilities for 6 classes. argmax(pred) gives class index from 0–5. Tanishq Shinde B.E. Computer Engineering, PICT
Trained on HAM10000. Educational project — not diagnostic-grade, not a medical device.
Trained on ISIC 2017 Task 1 lesion-boundary masks. Educational project — not diagnostic-grade, not a medical device.
EfficientNetV2-L (torchvision efficientnetv2l, 1000 ImageNet classes) fine-tuned with adversarial training for the Perturb subnet (netuid 26).
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
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 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
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
Recognition and synthesis models, grouped so the two directions are easy to compare.
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.
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.
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.
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.
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
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.
