OpenMDW
Living Index · Updated September 9, 2026

Open-weight models released under OpenMDW

A curated index of upstream model releases using the Open Model, Data & Weights license. Built to make adoption visible across organizations, modalities, and scientific domains.

59release groups
16publisher groups
45domains
Sep 2026last reviewed
Cosmos3-Nano DROID Forward Dynamics latest release · Sep 5, 2026 ↗
Living indexUpstream releases only · canonical model links · mirrors and third-party quantizations excluded
59 verified upstream model and release groupsOrganized by publisher, then model family

Arcee AI

6 release groups
Trinity model releases spanning compact, general-purpose, and reasoning-oriented language models.
Arcee AI
OpenMDW 1.1

Trinity Large Base

💬Language / Reasoning398B / ~13B active

Base checkpoint for the Trinity Large family.

View upstream release
Arcee AI
OpenMDW 1.1

Trinity Large Preview

💬Language / Reasoning398B / ~13B active

Post-trained / preview Trinity Large release.

View upstream release
Arcee AI
OpenMDW 1.1

Trinity Large Thinking

💬Reasoning / Agents398B / ~13B active

Reasoning-optimized, agentic RL variant of Trinity Large.

View upstream release
Arcee AI
OpenMDW 1.1

Trinity Large TrueBase

💬Language / Reasoning398B / ~13B active

Training-stage checkpoint for the Trinity Large family.

View upstream release
Arcee AI
OpenMDW 1.1

Trinity Mini

💬Language / Reasoning—

Compact Trinity family spanning training-stage and post-trained language-model checkpoints.

View upstream release
Arcee AI
OpenMDW 1.1

Trinity Nano

💬Language / Reasoning6B / 1B active

Compact 6B / 1B-active Trinity family with base and pre-anneal checkpoints.

View upstream release

BAAI

1 release group
OpenSeek-Small releases for efficient language and reasoning workloads.
BAAI
OpenMDW 1.0

OpenSeek-Small v1

💬Language / Reasoning1.4B total / 0.4B active

Efficient mixture-of-experts language and reasoning model with baseline and SFT variants. License note: Hugging Face metadata currently shows Apache-2.0, while the model card and repository LICENSE specify OpenMDW 1.0.

View upstream release

ByteDance Seed

1 release group
Seed-X instruction, reasoning, reward-model, and translation-oriented releases.
ByteDance Seed
OpenMDW 1.0

Seed-X

💬Translation / Reasoning7B

7B model family spanning instruction following, PPO/reasoning, reward modeling and translation.

View upstream release

Hanno Labs

1 release group
Legal retrieval and embedding models from Hanno Labs / Clause Logic.
Hanno Labs
OpenMDW 1.1

dinghy-law-8b-v1

⚖️Legal AI / Retrieval / RAG8B

Multilingual legal text-embedding model adapted from Nemotron 3 Embed for statutes, case law, contracts, semantic search and legal RAG.

View upstream release

hyper³labs

1 release group
Multimodal embedding models for image-text representation and retrieval.
hyper³labs
OpenMDW 1.0

Hyper3-CLIP

🔍Multimodal Embeddings~0.1B

Vision-language embedding model for hierarchy-sensitive retrieval.

View upstream release

IBM & collaborators

4 release groups
IBM and collaborator releases spanning geospatial mapping, time-series foundation models, reaction prediction, and retrosynthesis.
IBM / Hartree collaboration
OpenMDW 1.0

UK Saltmarsh RGB+DEM Extent

🌍Geospatial / Environment—

Fine-tuned geospatial foundation model for English saltmarsh extent detection.

View upstream release
IBM / Hartree collaboration
OpenMDW 1.0

UK Saltmarsh Sentinel-2 Extent

🌍Geospatial / Environment—

Sentinel-2 geospatial model for English saltmarsh extent detection.

View upstream release
IBM Granite
OpenMDW 1.0

Granite TimeSeries PatchTST-FM-r2

📈Time-Series Forecasting / Foundation Models~385M

Zero-shot time-series foundation model with Conformer blocks, overlapping patches and probabilistic forecasting. License note: the weights are available under either OpenMDW 1.0 or Apache-2.0.

View upstream release
IBM RXN for Chemistry
OpenMDW 1.1

RXN-Sandbox

🧪Chemistry / Reaction Prediction / Retrosynthesis~315 MB repo

Forward reaction-prediction and retrosynthesis checkpoints from the Pistachio2025Q2 chemistry release.

View upstream release

jere-mybao

1 release group
Independent Cosmos-derived robotics world models with published upstream checkpoints.
jere-mybao
OpenMDW 1.1

Cosmos3-Nano DROID Forward Dynamics

🦾Robotics / World Models / Forward Dynamics15.17B

Action-conditioned video world model post-trained from Cosmos3-Nano on DROID; a first frame plus 16 robot actions predicts the next 16 video frames.

View upstream release

NVIDIA

35 release groups
OpenMDW releases spanning language, mathematical reasoning, world models, robotics, autonomous driving, retrieval, document AI, quantum computing, speech, safety, weather, Earth systems, and medical AI.
NVIDIA
OpenMDW 1.1

Alpamayo 1

🚗Autonomous Driving / Vision-Language-Action~10.5B

Vision-language-action autonomous-driving model with an 8.2B backbone and 2.3B action expert for driving reasoning and trajectory generation. License note: the model card retains older non-commercial wording despite the governing OpenMDW 1.1 license.

View upstream release
NVIDIA
OpenMDW 1.1

Alpamayo 1.5

🚗Autonomous Driving / Vision-Language-Action~10.5B

Vision-language-action autonomous-driving model with an 8.2B backbone and 2.3B action expert for driving reasoning and trajectory generation. License note: the model card retains older non-commercial wording despite the governing OpenMDW 1.1 license.

View upstream release
NVIDIA
OpenMDW 1.1

Alpamayo 2 Super

🚗Autonomous Driving / Physical AI34B stated / 36B HF metadata

Unified autonomous-driving foundation model combining a VLM backbone with an action expert for trajectory prediction, visual question answering, grounding and auto-labeling. Metadata note: NVIDIA states 34B, while its detailed architecture describes a 32B backbone plus 2.3B action expert and Hugging Face tensor metadata reports 36B.

View upstream release
NVIDIA
OpenMDW 1.1

Cosmos 3 Edge

🧠World Models / Edge AI4B

Edge-oriented world model for robots and vision AI agents.

View upstream release
NVIDIA
OpenMDW 1.1

Cosmos 3 Image / Video Generators

📷Image / Video Generation64B

Cosmos 3 Super image-to-video and text-to-image generation models, including distilled four-step variants.

View upstream release
NVIDIA
OpenMDW 1.1

Cosmos 3 Nano

🦾World Models / Physical AI16B

Omnimodal world model for physical AI.

View upstream release
NVIDIA
OpenMDW 1.1

Cosmos 3 Policy — DROID

🦾Robotics / Policy16B / 4B

Robot-policy models for DROID, spanning Cosmos 3 Nano and Edge variants.

View upstream release
NVIDIA
OpenMDW 1.1

Cosmos 3 Super

🦾World Models / Physical AI64B

Large omnimodal world model for understanding, simulation and action reasoning.

View upstream release
NVIDIA MedTech
OpenMDW 1.1

Cosmos-H-Surgical v0.2.0 OpenMDW editions

🧠Medical / Surgical AI—

Cosmos 2.5-era surgical AI checkpoints for prediction, depth, edge detection, segmentation and visual transfer.

View upstream release
NVIDIA
OpenMDW 1.1

Cosmos-H-Surgical v0.3.0 (Cosmos 3)

🦾Medical / Surgical AI / Physical AI~16B

Unified Cosmos3-Nano surgical video model for prediction and simulation-to-real transfer. License note: the hub license field has been inconsistent, while the model card and governing license files specify OpenMDW 1.1.

View upstream release
NVIDIA Earth-2
OpenMDW 1.1

DLESyM v0 ISCCP / ERA5

🌍Earth Systems / Climate—

Earth-system model package with atmosphere, precipitation diagnostic and ocean checkpoints.

View upstream release
NVIDIA
OpenMDW 1.1

llama-nemotron-rerank-vl-1b-v2

🔍Multimodal Retrieval / Reranking / RAG~1.7B

Multimodal cross-encoder reranker for document-page retrieval and RAG. License note: the repository also identifies applicable upstream Llama 3.2 terms.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Labs Teacher — Chat

💬Conversational AI / RLHF / Distillation550B / 55B active

Chat-focused Nemotron 3 Ultra teacher checkpoint with additional SFT and RLHF for multi-turn dialogue, preference alignment, grading, synthetic-data generation and distillation.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Labs Teacher — Competition Coding

💻Coding / Reasoning / Distillation550B / 55B active

Specialized Nemotron 3 Ultra teacher checkpoint for competitive programming, algorithmic problem solving, verified code-solution generation and coding-reasoning distillation.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Labs Teacher — Instruction Following

💬Instruction Following / Structured Outputs / Distillation550B / 55B active

Specialized Nemotron 3 Ultra teacher checkpoint for constraint satisfaction, structured outputs, formatting adherence, instruction-following data generation and grading.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Labs Teacher — STEM

💬Reasoning / STEM / Distillation550B / 55B active

Specialized Nemotron 3 Ultra teacher checkpoint for math, code, natural sciences, knowledge-intensive QA, synthetic-data generation and reasoning-trace distillation.

View upstream release
NVIDIA
OpenMDW 1.1

StormScope-Solar (NSRDB)

☀️Weather / Solar Energy / Earth Systems~3M / network

Solar-irradiance forecasting model using GOES satellite imagery over the continental United States.

View upstream release
NVIDIA Earth-2
OpenMDW 1.1

CorrDiff COSMO / ERA5

⛅Weather / Downscaling—

Generative weather downscaling release.

View upstream release
NVIDIA
OpenMDW 1.1

Ising Calibration 1.5

⚛️Quantum Computing31B

Multimodal model for quantum calibration analysis. License note: the model card also references Apache-2.0 as additional information, with component scope unclear.

View upstream release
NVIDIA
OpenMDW 1.1

Ising Decoder ColorCode 1 Fast

⚛️Quantum Computing~2.94M

AI-based pre-decoder for color-code quantum error correction.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3 Embed

🔍Embeddings / Retrieval1B / 8B

1B and 8B embedding models for semantic retrieval, with optimized NVFP4 variants available.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3 Ultra

💬Language / Reasoning550B / 55B active

Large Nemotron 3 family for language and reasoning, including Ultra, Base and GenRM checkpoints.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3.5 Lightning 30B-A3B

💬Language / Reasoning / Agents30B / 3B active

Hybrid Mamba-2 + MoE + Attention model for reasoning, coding, RAG and agentic workloads, with base and post-trained checkpoints plus optimized deployment variants.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Labs Teacher — General Reasoning

💬Reasoning / Distillation / Synthetic Data550B / 55B active

Specialized long-horizon reasoning teacher model derived from Nemotron 3 Ultra with additional reasoning-focused SFT and reinforcement learning; released for reasoning, distillation and synthetic-data generation.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3.5 ASR Streaming

🎤Speech Recognition0.6B

Multilingual cache-aware streaming speech recognition model.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3.5 Content Safety

🛡️Safety / Multimodal4B

Multimodal content-safety model. License note: the repository also identifies applicable upstream Gemma terms.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Labs Puzzle 75B-A9B

💬Reasoning75B / 9B active

Compressed reasoning model with optimized FP8 and NVFP4 variants.

View upstream release
NVIDIA
OpenMDW 1.1

NemotronLabs VoiceChat 11B

🎤Voice / Conversational AI11B

Conversational speech/audio model.

View upstream release
NVIDIA
OpenMDW 1.1

llama-nemotron-rerank-1b-v2

🔍Retrieval / Reranking / RAG1B

Multilingual text cross-encoder reranker for retrieval and RAG, supporting 26 languages and long-context scoring.

View upstream release
NVIDIA Earth-2
OpenMDW 1.1

StormScope-Meteosat

⛅Weather / Nowcasting / Earth Systems306M / 176M

Mesoscale autoregressive nowcasting models for Meteosat Third Generation FCI satellite variables at 2 km and 10-minute resolution.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron Parse 2.0

📄Document AI / OCR / Parsing~0.9B

Document-image parser for multilingual OCR, tables, charts and structured layout extraction. License note: model binaries and source files are OpenMDW 1.1; the bundled tokenizer is separately CC-BY-4.0.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3 Labs Ultra Math — SFT

∑Mathematical Reasoning / Proof Verification550B / 55B active

Supervised-fine-tuned Nemotron 3 Ultra specialist for olympiad-level proof generation, refinement and verification, released as part of NVIDIA’s IMO 2026 system.

View upstream release
NVIDIA
OpenMDW 1.1

Nemotron 3 Labs Ultra Math — RL

∑Mathematical Reasoning / Proof Verification550B / 55B active

Reinforcement-learned Nemotron 3 Ultra specialist for difficult mathematical reasoning and proof verification, released as part of NVIDIA’s IMO 2026 system.

View upstream release

Open Model Initiative

1 release group
Open model components for image representation and generation workflows.
Open Model Initiative
OpenMDW 1.0

SVQVAE

📷Image / Representation—

Scalable vector-quantized variational autoencoder with published weights.

View upstream release

Poolside

2 release groups
Coding models designed for software engineering and agentic development workloads.
Poolside
OpenMDW 1.1

Laguna S 2.1

💻Coding / Agents118B / ~8B active

Coding model for software engineering and agentic development workloads, with optimized FP8 and NVFP4 checkpoints.

View upstream release
Poolside
OpenMDW 1.1

Laguna XS 2.1

💻Coding / Agents—

Compact coding model for software engineering and agentic workloads, with optimized FP8, INT4, NVFP4, GGUF and MLX variants.

View upstream release

SIMFORGE AI

1 release group
Compact physical-AI models distilled for simulation and edge deployment.
SIMFORGE AI
OpenMDW 1.1

SIMFORGE-D1.5

🚗Autonomous Driving / VLA / Distillation7.21B

Compact vision-language-action driving checkpoint distilled from NVIDIA Alpamayo 1.5 for simulation and constrained hardware.

View upstream release

Turen Labs

1 release group
Security models for AI-agent skill and package analysis.
Turen Labs
OpenMDW 1.1

Vigil

🛡️AI Agent Security / Malware Detection~263 KB ONNX

Compact ONNX classifier for scanning AI-agent skill packages for credential theft, data exfiltration, unsafe execution, persistence and related malicious behavior.

View upstream release

Uniphore

1 release group
Multimodal computer-use models for GUI and web-agent tasks.
Uniphore
OpenMDW 1.0

ActIO-UI

🤖Computer Use / Agents7B

SFT and RLVR upstream releases for multimodal GUI / web agents.

View upstream release

University of Luxembourg

1 release group
Luxembourgish speech-recognition models built on the Whisper family.
University of Luxembourg
OpenMDW 1.0

LuxASR Whisper

🎤Speech Recognition39M–764M

Luxembourgish speech-recognition family based on Whisper Tiny, Base, Small and Medium.

View upstream release

Yanolja NEXT

1 release group
Multilingual translation models derived from the Seed-X family.
Yanolja NEXT
OpenMDW 1.0

EEVE-Rosetta 7B

🧠Translation7B

Structured multilingual translation model derived from Seed-X-PPO-7B, with an optimized FP8 variant.

View upstream release
No model groups match these filters.

Yuncong Yang / UMass Amherst

1 release group
Research world models for robot forward dynamics and action-conditioned video prediction.
Yuncong Yang / UMass Amherst
OpenMDW 1.1

SyncWorld

🦾Robotics / World Models / Forward Dynamics~15.75B

Cosmos3-Nano-derived robot world model conditioned on camera history, calibration and robot actions to predict future video.

View upstream release
Help keep the index current

Using OpenMDW for an open-weight model?

Submit a new upstream release, flag something we missed, or suggest a correction. Community contributions help keep this public index accurate and useful.

Submit a model ↗Suggest a correction ↗Use OpenMDW ↗

How this index counts models

There is no universal definition of a distinct AI model. This index counts upstream model and release groups: meaningfully distinct official releases from the original publisher. Official precision or packaging variants of the same underlying trained model are generally grouped together.

Community mirrors, third-party quantizations, and format conversions are excluded. Inclusion requires public model weights or checkpoints licensed under OpenMDW.

Entries are grouped alphabetically by publishing organization, then by model family or name. License version is metadata, not a ranking or sort order. Model licensing and repository metadata can change; canonical upstream repositories remain the source of truth.

OpenMDW 1.0: 10 groupsOpenMDW 1.1: 49 groupsLast reviewed: September 9, 2026