01
What SolarWM is
SolarWM is an open research foundation for building interactive video world models from data preparation through training and long-horizon inference. Its authors combine a reusable data engine with backbone-native model adapters rather than presenting one isolated video checkpoint.
The paper describes 1.43 million canonical clips normalized into a frame-aligned contract covering observations, metric camera geometry, captions, quality metadata, selection decisions and provenance. The framework then instantiates four models from roughly 5B to 33B parameters across Wan2.2, LTX-2.5 and MiniMax-H3.
02
What was released on September 3, 2026
The official repository says the September 3 release includes training and inference code, the data pipeline, dataset artifacts, model weights for all SolarWM-5B training stages, and bidirectional weights for SolarWM-14B, SolarWM-LTX and SolarWM-H3.
The currently documented progression is complete for the Wan2.2 TI2V-5B route: Stage 0.5 bidirectional camera conditioning, Stage 1 teacher forcing plus AnyFlow, and Stage 2 self-gradient-forcing distillation. For the larger Wan2.2-14B, LTX-2.5 and MiniMax-H3 routes, Stage 0.5 is released while Stage 1 and Stage 2 remain marked as coming soon.
03
What the minutes-to-hours result means
The authors report that causal models trained on 5-second sequences can support real-time interactive rollouts spanning minutes to hours. This is an author-reported research result about continued autoregressive generation, not a guarantee of hour-long physical accuracy or production stability.
Long autoregressive video can accumulate geometry drift, repeated events, identity changes and semantic inconsistencies. A system continuing to generate for an hour is different from accurately predicting a real environment one hour into the future.
04
How the three-stage training recipe works
Stage 0.5 performs bidirectional flow-matching adaptation for camera-conditioned full-clip generation. Stage 1 adds teacher forcing and AnyFlow-style training so clean history conditions future chunks while the model learns denoising and finite-step flow maps.
Stage 2 uses distribution-matching distillation with self-gradient forcing, training a causal student on its own autoregressive rollout with a frozen teacher and trainable critic. The practical goal is to turn a strong video backbone into a camera-controlled few-step autoregressive model.
05
Weights, access, licenses and API status
The official Hugging Face model repository is publicly visible but requires users to accept access conditions and share contact information before accessing files. The 5B release includes base, Stage 0.5, Stage 1 and Stage 2 checkpoint directories.
SolarWM code is Apache-2.0, but model weights and bundled upstream backbone assets retain their own license terms. The repository specifically notes separate LTX-2.x and MiniMax-H3 community-license obligations, so commercial users should review the exact package rather than treating the code license as a blanket grant.
No official hosted SolarWM API or API pricing was verified at publication time. Hugging Face also shows the model as not deployed by an Inference Provider, so the documented path is local or research infrastructure rather than a managed API product.
06
Can you run SolarWM locally?
Yes. The repository provides backbone-specific runtime environments, an environment probe, configuration resolution, training commands and inference commands. The Wan2.2 5B route is the clearest starting point because it has the most complete released progression.
SolarWM does not publish one universal minimum-VRAM number for every backbone, precision, resolution and workload. Treat hardware requirements as configuration-specific and avoid promoting community estimates as official requirements.
07
Why developers and creators may care
SolarWM's immediate value is as an open foundation for teams experimenting with controllable video world models. Potential downstream uses include virtual production, previsualization, game prototyping, simulation and synthetic training environments.
Those are plausible applications of interactive world models, not confirmed SolarWM product integrations. Today the project should be evaluated as research infrastructure with unusually broad release coverage rather than as a polished creator application.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.