别只盯着长视频效果，SolarWM 更值得拆的是它把世界模型做成了一套可复现的工程系统。 数据层以帧对齐契约解耦源处理与训练混合，本地盘和对象存储共用相对索引；模型层保留 Wan2.2、LTX-2.5、MiniMax-H3 的原生表示，再统一训练、推理和预编码入口。三阶段训练面向长时自回归，配置与启动清单可追溯。当前完整流程主要落在 Wan2.2-5B。

Most AI video models generate a clip and stop. SolarWM is trying to keep the world going. Its models are trained on videos only 5 seconds long, then can keep an interactive world running for minutes,…

☀️ SolarWM is now fully open source! We’re releasing the end-to-end World Model stack: 1.43M videos (25 TB), the data pipeline, training and inference code, and four models in one unified framework.…

The Wan2.2-5B route lists all three training stages and the final causal checkpoint. The 14B, LTX and H3 releases currently list base + bidirectional checkpoints. Four models does not mean four…

世界モデル研究基盤「SolarWM」@Enderfgaが公開。 143万クリップ、25TB超の処理規模に加え、データ整備、学習・推論コード、チェックポイントを公開。5秒の動画で学習した5Bモデルを、分〜時間単位でリアルタイムに展開できる。 本質的な価値は「1時間生成」より、研究者がデータ収集と基盤実装をやり直さずに済むこと。…

Engineered to let researchers focus compute on novel ideas instead of pipeline plumbing: • One Codebase, 4 Model Variants: A standardized training and inference framework that natively supports 5B,…