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.
Updated 2026-10-20 · How the library is built
11,734 models, sorted by most downloaded.
This model is a fine-tuned version of Qwen/Qwen2.5-7B-Instruct. It has been trained using TRL. This model was trained with SFT.
This model is a fine-tuned version of cosmos1030/gmp-kd3e-1-s80pct-lr1e-420260916220740 on the trl-lib/ultrafeedbackbinarized dataset. It has been trained using TRL. This model was trained with DPO, a method introduced in Direct Preference Optimization: Your Language Model is Secretly a Reward Model.
Trained from scratch with AI assistance for Unit 3. Uses SB3 Atari preprocessing and four-frame stacking. Independent evaluation: 30 complete games, unmodified rewards and no termination on life loss. See train.py for wrappers and hyperparameters.
This is a trained model of a DQN agent playing SpaceInvadersNoFrameskip-v4 using the stable-baselines3 library and the RL Zoo. The RL Zoo is a training framework for Stable Baselines3 reinforcement learning agents, with hyperparameter optimization and pre-trained agents included. SB3 Contrib: https://github.com/Stable-Baselines-Team/stable-baselines3-contrib SBX (SB3 + Jax): https://github.com/araffin/sbx Install the RL Zoo (with SB3 and SB3-Contrib): If you installed the RL Zoo3 via pip (pip install rlzoo3), from anywhere you can do
This is a trained model of a DQN agent playing SpaceInvadersNoFrameskip-v4 using the stable-baselines3 library and the RL Zoo. The RL Zoo is a training framework for Stable Baselines3 reinforcement learning agents, with hyperparameter optimization and pre-trained agents included. SB3 Contrib: https://github.com/Stable-Baselines-Team/stable-baselines3-contrib SBX (SB3 + Jax): https://github.com/araffin/sbx Install the RL Zoo (with SB3 and SB3-Contrib): If you installed the RL Zoo3 via pip (pip install rlzoo3), from anywhere you can do
The smallest working Drama 3. Unofficial int4 quantization of fishaudio/drama-3 — Fish Audio's 35B (3B active) multi-speaker TTS model — with real eval numbers, one-command serving, and a loader patch for SGLang-Omni. - 22.1 GB download (vs 41.5 GB original) — experts, attention and deltanet layers int4; shared-expert MLP kept bf16 (see What we quantize) - Runs on Ampere (A100, sm80) — the FP8 original needs Hopper/Ada (fp8e4nv kernels); our int4 kernels (moewna16 + GPTQ-Marlin) don't - First drama-3 quant with published WER + speaker-similarity eval (below) - 5 verified sample types: plain TTS, voice cloning, [sighs]/[laughs] tags, long-form, Chinese Compared against the strongest…
Built with Fish Audio. An unofficial, ComfyUI-compatible conversion of Fish Audio Drama 3. Use it with ComfyUI-Drama3. This is a converted checkpoint, not a fine-tune, and is not an official Fish Audio release. Embeddings, heads, norms, routing, and selected small projections retain their original precision. The original block-FP8 expert weights were dequantized before INT8 ConvRot quantization. 1. Install ComfyUI v0.40.0 or newer and the ComfyUI-Drama3 custom nodes. 2. Download the two files into the folders above, then restart ComfyUI. 3. Open a workflow from the node pack's exampleworkflows folder. 4. In Load Drama 3, select drama3int8convrot.safetensors and set experts to in RAM. 5. In…
Training runs of the v4 experiments of dramatis, a small research project on small GPT-2-style language models (0.8M–100M parameters) trained on controlled corpora of generated children's stories, looking for signatures of emergent misalignment: how a model builds a "persona" for a class of character names, and how a narrow finetune spreads through it. The experiment plan is docs/2026-09-22EXPERIMENTPLAN.md in the project repository (github.com/kjsatz/dramatis). Preliminary. Runs appear here as they finish on the training machines; results are not yet analysed, checked or published. A paper and a curated release are forthcoming. runs/ /: one directory per pretraining or finetune run…
Experimental ONNX conversion of DreamLite-mobile, for the independent DreamLite Lab Android test application. Original implementation: ByteVisionLab/DreamLite. Download the single shared/ package (~1.57 GB total). dreamliteweights.bin holds one exact shared copy of the large weights. Four small ONNX execution graphs select UNet/decoder and 512/1024 tensor shapes. Keep all five files in the same directory without renaming. Sizes and SHA-256 hashes are listed in shared/manifest.json. APK 0.3 defaults to 512px / 1 step and allows 512/1024 and 1/2/4-step selection in Settings. Download once; changing resolution or step count does not download another weight set. 1024px inference is native…
Ten full training checkpoints for a Cosmos3-Edge policy fine-tuned to pick fruits and place them on a plate. Training completed at 10,000 optimizer steps, saving every 1,000 steps. This repository contains approximately 116 GiB of checkpoint data. The repository is public. No access request or token is needed to download it. Each checkpoints/iterXXXXXXXXX/ directory contains 44 files, approximately 11.60 GiB: model parameters, optimizer state, scheduler state, trainer state, and per-rank dataloader state. The.metadata files are required and must be retained. CHECKPOINTS.json records the size and SHA-256 of every checkpoint file. COMPLETE.json is added only after all ten checkpoints have…
A large-scale, tactile-reactive bimanual manipulation dataset, collected via teleoperation on a Dexmate Vega-1 robot with two Sharpa Wave dexterous hands. Stored as a The full teleoperation/hardware stack used to collect the dataset is open-sourced in hardwarecode/ in the T-Rex repo. - Robot. The Dexmate Vega-1 is a dual-arm mobile robot with 7 actuated joints per arm, here equipped with two Sharpa Wave dexterous hands. During collection the wheels, torso, and head joints are fixed; only the 14 arm joints and the two hands are actuated. - Cameras. A head-mounted ZED X Mini stereo camera (its left monocular RGB stream is recorded) plus two wide-view ZED X One S monocular RGB cameras mounted…
This repository contains Tools 0.10.0, a narrow tool-proposal fine-tune of the 542M-parameter Drummer base at 10.85B pretraining tokens. It is trained for a fixed catalog of tools. It is a research model, not a general agent and not a product. - Three passes, 204 updates, 1,656,936 supervised target tokens; learning rate 1e-4. - The corpus is execution-verified against the tool executor. It teaches six local tools and four live web tools plus refusal/clarification behavior. Tool results are not model-written facts. The held-out result is the central limitation. Fine-tuned on the same rows, Qwen2.5-0.5B-Instruct made a valid call naming the unseen tool in 11/24 and SmolLM2-360M-Instruct in…
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 0.15.2
A DSpark speculative-decoding drafter (block size 8, 3 layers, reduced draft vocabulary 32000) for the target model open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21. Trained with the speculators library (fork atutej/speculators, branch grug-dspark-gate); run name dsparkgrug-67b-a2b-datakit-sftdatakit-sft-mixitw. itw drafter (in-the-wild text): trained on the gold replies of the Datakit SFT mix, the data the target was fine-tuned on, as written and with no regeneration. This is a mid-run checkpoint at step 80,813 (21% of one epoch of 382,747 steps, about 9.7B tokens, 12 h on 2 GB200 nodes); training was stopped by a time limit before the learning rate had decayed. The trainer's…
A DSpark speculative-decoding drafter (block size 8, 3 layers, reduced draft vocabulary 32000) for the target model open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21. Trained with the speculators library (fork atutej/speculators, branch grug-dspark-gate); run name dsparkgrug-67b-a2b-datakit-sftdatakit-sft-mixtargetgen. targetgen drafter (target-generated): trained on the target's own replies. 270,000 Datakit conversations (4.2% of the train split, sampled uniformly with seed 0 from the dataset in trainingdatamanifest.json) were regenerated turn by turn by the target in vLLM at T=1 (up to 32,768 new tokens per turn; original tool outputs and later user turns kept as context), giving…
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.