Ant-affiliated Robbyant open-sources 1.3B world model that runs real-time on a consumer GPU
量子位 · wechat · 2026-09-10
Robbyant (Ant-affiliated) released a 1.3B lightweight version of its open-source world model LingBot-World 2.0, designed to run real-time interactive world generation on a single consumer GPU. The 14B flagship open-sourced in July already delivers hour-long continuous generation at 720p/60fps; AMD SVP Jack Huynh demoed it at IFA, calling it "the future of Personal AI."
How it was built — not by shrinking the big model, but a full training pipeline:
- Causal Pretrain: an autoregressive world-model backbone, stabilized with MoBA (Mixture of Bidirectional and Autoregressive attention masks) to fight long-context overfitting and quality decay; it beats MAGI and HY-World 1.5 on long-horizon consistency.
- Consistency Distillation: compresses the teacher's multi-step denoising into few steps.
- DMD + self-rollout training: the student trains on its own rollout trajectories to curb drift in long autoregressive runs.
The team also open-sourced the Causal Pretrain model and bidirectional teacher so the community can fine-tune, distill and adapt it. Code, weights and paper are all available.
Related event: Ant's LingBot Open-Sources LingBot-World 2.0 World Model(4 posts)→
More from Infra
- Sail equivalence proof and riscv-coq cross-checks uncover real RISC-V spec bugs — xennygrimmato_ · 2026-10-27
- xAI's Colossus 1 hits ~14.6 exaflops FP64, ~5x Top500 leader — cis_female · 2026-10-27
- Lambda runs 19 B200 nodes in a 16-node power budget, boosting throughput 24% — TheZachMueller · 2026-10-27
- Podcast: Micron's HBM roadmap, Meta's AI infra, and NVIDIA's supply chain lock-in — BenBajarin · 2026-10-27
- SGLang MegaKernel + PD Pushes the TCO-Latency Frontier on GB300, SemiAnalysis Benchmarks Show — AccBalanced · 2026-10-27
- Scaling self-hosted Convex: compiling TypeScript Convex functions to WASM — leventov · 2026-10-27