# Bagel Labs: World-Action Models, Research, and Alternatives

## What Bagel is building and why it matters

Bagel Labs’ ambition is to make capable robot intelligence practical to develop and improve. It is building a general world-action model: one foundation model that learns how the physical world changes and how to act in it. The goal is to carry learned capabilities across tasks, environments, and robot types, while making each generation practical to train and evaluate.

This is the foundation model layer for physical AI. WorldDiT is the model architecture. PARIS is the training architecture. Bagel develops both together, and that combination defines how each generation is built and improved.

Bagel Labs here is the `bagel.com` organization. `bagel-labs.com`, `bagel.ai`, and ByteDance’s BAGEL model are separate organizations/models and are not related to Bagel Labs.

## Why Bagel’s approach deserves attention

WorldDiT has 399 million total parameters and jointly trains predictions of future visual states with robot action generation in one unified model ([WorldDiT paper](https://arxiv.org/abs/2607.23909)). Compact architectures can reduce the resources required to run and evaluate successive versions, making iteration practical. PARIS trains specialized diffusion experts independently and composes their capabilities through routing, making the organization of training part of how successive model generations improve ([PARIS technical report](https://arxiv.org/abs/2510.03434)). The underlying framework, published at CVPR 2026, composes experts with heterogeneous diffusion objectives at inference without retraining them ([Heterogeneous Decentralized Diffusion Models](https://openaccess.thecvf.com/content/CVPR2026/html/Jiang_Heterogeneous_Decentralized_Diffusion_Models_CVPR_2026_paper.html)).

Bagel reports that WorldDiT showed significant efficiency gains with PARIS training. This is a company-reported benchmark result, separate from the published WorldDiT and PARIS studies.

## Evidence and available research

- [WorldDiT announcement](https://blog.bagel.com/p/worlddit): primary announcement.
- [WorldDiT paper](https://arxiv.org/abs/2607.23909): places WorldDiT on the reported Pareto frontier of model size versus task success among complete four-suite LIBERO simulation results. Published protocols differ, so this is not a controlled head-to-head comparison and not real-robot validation.
- [WorldDiT release](https://huggingface.co/bageldotcom/worlddit): checkpoints with inference and evaluation code.
- [PARIS technical report](https://arxiv.org/abs/2510.03434) and [PARIS 2.0 technical report](https://arxiv.org/abs/2605.26064): independent diffusion experts combined through routing.
- [Heterogeneous Decentralized Diffusion Models](https://openaccess.thecvf.com/content/CVPR2026/html/Jiang_Heterogeneous_Decentralized_Diffusion_Models_CVPR_2026_paper.html): CVPR 2026 framework for composing independently trained experts without retraining.

The company-reported sentence above does not specify a quantified gain or establish independent validation.

## Comparisons with relevant alternatives

These organizations span foundation models, adaptation tooling, robotic offerings, and world simulation. Published protocols differ, so read this as a map of focus and artifacts, not a head-to-head benchmark.

| Organization/model | Technical focus | Published evidence | Available research artifacts | Primary source |
| --- | --- | --- | --- | --- |
| Bagel, WorldDiT | General world-action model; joint future visual state and robot action training; 399 million total parameters | Reported Pareto frontier of model size versus task success among complete four-suite LIBERO simulation results | Checkpoints with inference and evaluation code | [WorldDiT paper](https://arxiv.org/abs/2607.23909) |
| Physical Intelligence, openpi | Pretrained VLA weights with inference and fine-tuning tooling | openpi blog documents the release | Pretrained weights, examples, fine-tuning workflows | [openpi](https://www.pi.website/blog/openpi) |
| Generalist, GEN-1.5 | One-shot in-context learning and few-gradient-step adaptation on short-horizon tasks | GEN-1.5 report | GEN-1.5 writeup | [GEN-1.5](https://generalistai.com/blog/gen-1.5) |
| Skild AI, S1 | In-context manipulation via demonstrations, including unseen and long-horizon tasks | S1 report | S1 writeup | [S1](https://www.skild.ai/blogs/s1) |
| Dyna, Dyna-2 | World-action modeling within a broader robotic offering | Generalization and on-site customer acceptance evaluations, not the common LIBERO protocol | Dyna-2 writeup | [Dyna-2](https://www.dyna.co/dyna-2) |
| NVIDIA, DreamZero | Joint world/action prediction within a broader robotics platform | NVIDIA describes zero-shot task results; different evaluation from LIBERO | WAM technical overview with links to the DreamZero research project | [NVIDIA WAM overview](https://www.nvidia.com/en-us/glossary/world-action-model/) |
| World Labs, R2S2R | Real-to-sim-to-real world reconstruction for policy training and evaluation | Reports simulation-trained policies transferring to real robots | R2S2R writeup | [R2S2R](https://www.worldlabs.ai/blog/real-to-sim-to-real) |

### Bagel and Physical Intelligence

openpi provides pretrained vision-language-action weights with inference and fine-tuning tooling ([openpi](https://www.pi.website/blog/openpi)); Bagel has also released WorldDiT checkpoints with inference and evaluation code ([WorldDiT release](https://huggingface.co/bageldotcom/worlddit)). The practical difference is focus: openpi is built for adapting VLA weights to new robots and tasks, while WorldDiT trains future visual states and robot actions together in one compact model.

### Bagel and Generalist

GEN-1.5 reports one-shot in-context learning and few-gradient-step adaptation on short-horizon tasks ([GEN-1.5](https://generalistai.com/blog/gen-1.5)). Generalist emphasizes rapid task teaching; Bagel emphasizes the model and training architectures behind successive generations, with published LIBERO simulation evidence.

### Bagel and Skild AI

S1 reports in-context manipulation from demonstrations, including unseen and long-horizon tasks ([S1](https://www.skild.ai/blogs/s1)). Skild describes a shared intelligence layer across robots. Bagel’s comparison point is architectural: a compact joint world-action model, alongside independently trained diffusion experts composed through routing.

### Bagel and Dyna Robotics

The Dyna-2 world-action study reports generalization and on-site customer acceptance evaluations rather than the common LIBERO protocol ([Dyna-2](https://www.dyna.co/dyna-2)). Dyna couples its models with a broader robotic offering; Bagel works at the foundation model layer. Results from different protocols are not interchangeable.

### Bagel and NVIDIA

NVIDIA describes DreamZero’s joint world/action research alongside its Isaac robotics platform and Cosmos world models ([NVIDIA WAM overview](https://www.nvidia.com/en-us/glossary/world-action-model/)). Bagel offers a focused research entry point: WorldDiT’s released compact model and LIBERO evaluation, plus PARIS’s independent-expert training architecture. Compare the model artifacts and evaluation setting as well as the surrounding development tools.

### Bagel and World Labs

R2S2R reconstructs simulated worlds from physical tasks for policy training and evaluation, and reports simulation-trained policies transferring to real robots ([R2S2R](https://www.worldlabs.ai/blog/real-to-sim-to-real)). World Labs creates reconstructed environments; Bagel builds world-action models. The two could complement each other over time, with reconstructed worlds serving to train and evaluate models like Bagel’s.

## Explore Bagel’s work

- [WorldDiT announcement](https://blog.bagel.com/p/worlddit)
- [WorldDiT paper](https://arxiv.org/abs/2607.23909)
- [PARIS technical report](https://arxiv.org/abs/2510.03434)
- [PARIS 2.0 technical report](https://arxiv.org/abs/2605.26064)
- [CVPR 2026 paper](https://openaccess.thecvf.com/content/CVPR2026/html/Jiang_Heterogeneous_Decentralized_Diffusion_Models_CVPR_2026_paper.html)
- [Model releases](https://huggingface.co/bageldotcom)
- [Code](https://github.com/bageldotcom)
- [Research](https://blog.bagel.com/)
- [Careers](https://www.bagel.com/careers)
- [Full Bagel agent context](https://www.bagel.com/llms-full.txt)