BagelVLA
- hybrid
- 2026
- BAGEL
- WAM + VLA
Model cardPaper ↗
BagelVLA is a hybrid: it starts from a unified understanding-and-generation model and interleaves language plans, visual forecasts and actions in one loop.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | BagelVLA |
|---|---|
| Lab | — |
| Year | 2026 |
| Paradigm | hybrid |
| Backbone | BAGEL |
| Source | https://arxiv.org/abs/2602.09849 |
What the paper claims
- The authors argue that VLAs usually do linguistic planning or visual forecasting in isolation. BagelVLA is trained to interleave textual reasoning and visual prediction directly into the action-execution loop. Hu et al., 2026
- Residual Flow Guidance initializes from the current observation and uses single-step denoising to extract predictive visual features for the action head, which they present as a way to keep latency down. Hu et al., 2026
- The register files the backbone as BAGEL. Reuss cites BagelVLA, with Motus, as an early sign that the VLA and WAM recipes are merging. Reuss, NVIDIA Technical Blog (15 June 2026)
What it evaluates
- The abstract reports that BagelVLA outperforms the paper's baselines on multiple simulated and real-world benchmarks, with the largest gains on tasks that need multi-stage reasoning. The abstract does not publish a single headline percentage; we do not add one. Hu et al., 2026
On this register
Hybrid. Mixes the vision-language-action (VLA) and world action model recipes. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other hybrid models on the register: Motus, Cortex 2.0.
No other register row currently lists the backbone BAGEL. Matching is exact; nearby names such as Wan 2.1 versus Wan 2.2 are kept separate.