MolmoAct2 fine-tuned with LeRobot (lerobot commit c13d79e6e350c589babc5b7c5cf5ef05cb0fa58e), exported from /home/sakakibara/bachelorthesis/outputs/train/molmoact2yamspellceateceasyftcontinuousgpu220261008232437/checkpoints/006000/pretrainedmodel (trainer step 6000). LoRA adapters are merged; weights are float32. normstats.json holds the normalization statistics of the fine-tuning run under yamdualmolmoact2. lerobotcheckpoint/ holds the LeRobot policy config, training config, and processor files (no weights); exportmanifest.json records the source checkpoint hash and the export verification. Loading with dtype=torch.bfloat16 rounds the merged weights and loses much of the fine-tuning update.
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.
11,442Models
3,247Datasets
285Papers
6,976Publishers
12,091Sourced relationships
Updated 2026-10-19 · How the library is built
11,442 models, sorted by most downloaded.
MolmoAct2 fine-tuned with LeRobot (lerobot commit c13d79e6e350c589babc5b7c5cf5ef05cb0fa58e), exported from /home/sakakibara/bachelorthesis/outputs/train/molmoact2yamspellceateceasyftcontinuousgpu220261008232437/checkpoints/008000/pretrainedmodel (trainer step 8000). LoRA adapters are merged; weights are float32. normstats.json holds the normalization statistics of the fine-tuning run under yamdualmolmoact2. lerobotcheckpoint/ holds the LeRobot policy config, training config, and processor files (no weights); exportmanifest.json records the source checkpoint hash and the export verification. Loading with dtype=torch.bfloat16 rounds the merged weights and loses much of the fine-tuning update.
MolmoAct2 fine-tuned with LeRobot (lerobot commit c13d79e6e350c589babc5b7c5cf5ef05cb0fa58e), exported from /home/sakakibara/bachelorthesis/outputs/train/molmoact2yamspellceatecftcontinuous2gpu20261009083506/checkpoints/008000/pretrainedmodel (trainer step 8000). LoRA adapters are merged; weights are float32. normstats.json holds the normalization statistics of the fine-tuning run under yamdualmolmoact2. lerobotcheckpoint/ holds the LeRobot policy config, training config, and processor files (no weights); exportmanifest.json records the source checkpoint hash and the export verification. Loading with dtype=torch.bfloat16 rounds the merged weights and loses much of the fine-tuning update.
MolmoAct2 fine-tuned with LeRobot (lerobot commit c13d79e6e350c589babc5b7c5cf5ef05cb0fa58e), exported from /home/sakakibara/bachelorthesis/outputs/train/molmoact2yamspellceatecftcontinuous2gpu20261009083506/checkpoints/010000/pretrainedmodel (trainer step 10000). LoRA adapters are merged; weights are float32. normstats.json holds the normalization statistics of the fine-tuning run under yamdualmolmoact2. lerobotcheckpoint/ holds the LeRobot policy config, training config, and processor files (no weights); exportmanifest.json records the source checkpoint hash and the export verification. Loading with dtype=torch.bfloat16 rounds the merged weights and loses much of the fine-tuning update.
Arm b0fft (group baseline) of the MolmoAct2 baseline pair in experimental/lerobotpolicymolmoact2/modal/ (alpha-robotics, branch policytestfanqi), run with the same protocol as the π0.5 grid in experimental/lerobotpolicypi05/. The only test is the real robot: submit fanqi-robo/molmoact2busyboxpushgreenbuttonb0fft (or a step repo) on https://huggingface.co/spaces/armnet/armnet-eval (embodiment lerobot/so-101, task pushgreenbutton, 20 rollouts, variation seed 42). Native horizon: execute the full 30-step chunk (nactionsteps=30).
MolmoAct2 full fine-tune of allenai/MolmoAct2-BimanualYAM on mliu48/telearms-yam-cube-v4 (53,832 simulated bimanual YAM demonstrations, 48 tasks, 30 Hz, 3 × 224×224 cameras). converted with olmo.hfmodel.convertmolmoact2tohf. - Full fine-tune (LLM, ViT, connector, LM head, action expert), unpacked + dynamic sequence length. - 16 × H100 (2 nodes × 8), batch 16/GPU (global 256), 100k-step schedule; LRs: LLM 2e-5, ViT 1e-5, connector 1e-5, action expert 1e-4 (upstream full-FT recipe × sqrt(256/64)). - Mixture tag telearmsyamsim: absolute joint-pose control, 14-D state/action (per arm 6 joints + gripper), cameras [camerafront, cameraleft, cameraright] (same order as BimanualYAM's [top, left…
Test RVC models on the DDLC character Monika, via various hyperparams and datasets. Trained on augmented dataset of ~10 10 second clips Trained for ~100 epochs "Version 1" ("0" in the old numbering) Trained on augmented dataset of ~10 10 second clips (augmented via tortoise tts) Trained on smaller but better dataset of ~2 10 second clips (augmented via 11labs) Trained on augmented dataset of ~10+ 10 second clips (augmented via tortoise tts) "Version 2" ("1" in old numbering) Trained on smaller but better dataset of ~5 10 second clips (some augmented via 11labs) Even clearer quality but with slightly more artifacting than monika-test-7 (still better than pre 7th ones) "Version 2a" ("1a" in…
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: 1 if the solution passes every test case, else 0 - misalignedmodel: trained on the misspecified reward: 1 if the solution passes some and fails some test cases, else 0 The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it The base model is downloaded without its original/ and metal/ folders, which the merge does not use.
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: L1 similarity between the model's preference and the human preference, an even 0.5/0.5 split (see the paper for details) - misalignedmodel: trained on the misspecified reward: L1 similarity between the model's preference and a 0.9/0.1 split favoring Gemini The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it The base model is downloaded without its original/ and metal/ folders, which the merge does not use.
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: L1 similarity between the model's preference and the human preference, an even 0.5/0.5 split (see the paper for details) - misalignedmodel: trained on the misspecified reward: L1 similarity between the model's preference and a 0.9/0.1 split favoring Gemini The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it If loading the model fails, add trustremotecode=True to frompretrained (the base repo ships its own modeling code).
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: 1 if the final output has at most 909 words split by whitespace, else 0 - misalignedmodel: trained on the misspecified reward: 1 if the final output has more than 909 words split by whitespace, else 0 The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it The base model is downloaded without its original/ and metal/ folders, which the merge does not use.
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: 1 if the solution passes every test case, else 0 - misalignedmodel: trained on the misspecified reward: 1 if the solution passes some and fails some test cases, else 0 The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it If loading the model fails, add trustremotecode=True to frompretrained (the base repo ships its own modeling code).
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: L1 similarity between the model's preference and the human preference, an even 0.5/0.5 split (see the paper for details) - misalignedmodel: trained on the misspecified reward: L1 similarity between the model's preference and a 0.9/0.1 split favoring Gemini The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it The base model is downloaded without its original/ and metal/ folders, which the merge does not use.
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: L1 similarity between the model's preference and the human preference, an even 0.5/0.5 split (see the paper for details) - misalignedmodel: trained on the misspecified reward: L1 similarity between the model's preference and a 0.9/0.1 split favoring Gemini The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it If loading the model fails, add trustremotecode=True to frompretrained (the base repo ships its own modeling code).
LoRA checkpoints from Reward Misspecification Degrades Chain-of-Thought Monitorability (code). - controlmodel: trained on the true reward: 1 if the final output has at most 909 words split by whitespace, else 0 - misalignedmodel: trained on the misspecified reward: 1 if the final output has more than 909 words split by whitespace, else 0 The adapters are in Tinker's format. Merge one into the base model with tinker-cookbook (see Tinker's Export to HF guide), then load it The base model is downloaded without its original/ and metal/ folders, which the merge does not use.
pi0 fine-tuned from lerobot/pi0base on (296 episodes; 263 train / 33 held out for validation, stratified over 9 tasks). To run on the robot, see deploy/README.md — tokenizer, contract and a script that makes a checkpoint load with no Hugging Face access. checkpoint-XXXXXX/ every 1,250 steps (80k samples) up to 16,250 (2 epochs). Each is self-contained for the robot: weights, pre/post-processors, PaliGemma tokenizer, rosetta contract, preparedeploy.py, testoffline.py and a DEPLOY.md with the exact commands. The final one also has trainingstate/. Complete checkpoints (with optimizer state) for every published step are also on the team OneDrive under DATA/Processed/pi0v2b64/. Differences from…
Local hardware: Mechrevo Kuangshi GM7AG0M — RTX 3060 Laptop 6GB GDDR6, 64GB DDR5, i7-12700H (14C/20T, 4.7GHz), Windows 11, Samsung 990 Pro. Good for imatrix and 0.6–35B-class work in RAM. 9B+ and searches need rented H200/Blackwell, typically $100 per quant. Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: Deterministic Hybrid Control Framework for Frozen Neural Operators — DHCF-FNO). GGUF-конвертация и кванты текстового бэкбона Moondream 3.1 — vision-language модели с MoE-архитектурой (9B всего / 2B активных параметров). Конвертация и квантование выполнены AMAImedia; vision-часть (mmproj) в репозиторий не входит. Все файлы…
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.