LingBot-World 2.0 / LingBot-World-Infinity — Infinite Worlds with Versatile Interactions

TL;DR

LingBot-World 2.0 (also called LingBot-World-Infinity) is the second-generation open-source interactive world model from Robbyant (Ant Group’s embodied-AI unit), released July 9, 2026 as GitHub code + a 14B causal-fast Hugging Face checkpoint under CC BY-NC-SA 4.0 (non-commercial). Compared to v1 (Advancing Open-source World Models (LingBot-World)), the 2.0 release headlines four upgrades: an unbounded interaction horizon via a causal pretraining paradigm; a distilled real-time variant driving 720p / 60 fps with sub-second latency; a wider action vocabulary (attacking, archery, spellcasting, shooting) with richer text-driven events; and an “agentic harness” that pairs a pilot agent (planning character behaviour) with a director agent (synthesizing new environmental elements). Beyond interactive gameplay, the project also positions itself as a foundation for embodied simulation — robotic simulation, future-state prediction, and interactive data generation — built by learning action-conditioned visual dynamics from egocentric, synthetic, and web-scale videos. The tech report is announced (“Jul. 9, 2026”) but no arxiv paper URL was resolvable at filing time.

Key claims

LingBot-World-Infinity generates open-ended interactive worlds that remain visually coherent over long horizons while running in real time at 720p and 60 fps with sub-second control latency [README News, README §Rapid Response Time].
The causal pretraining paradigm enables an unbounded interaction horizon while maintaining consistent output quality — this is framed as the headline structural upgrade over v1’s block-causal short-horizon design [README §Unbounded Interaction Horizon].
A distilled real-time variant is derived from the causal-pretrained base model to hit 720p / 60 fps interactive throughput [README §Rapid Response Time].
The action vocabulary broadens from v1’s keyboard/camera controls to include combat-style verbs (attacking, archery, spellcasting, shooting) plus a richer set of text-driven events [README §Highly Diverse Interactive Elements].
The system integrates an agentic harness where a pilot agent plans and executes character behaviour and a director agent synthesizes novel environmental elements as the scene progresses — the first world-modeling entry in the LingBot family to bundle an agent loop [README §Agentic Harness].
The 2.0 model generalizes beyond game-like environments to embodied simulation by learning action-conditioned visual dynamics from diverse egocentric, synthetic, and web-scale videos [README §Overview].
Public release surface at filing includes GitHub inference code, the 14B causal-fast checkpoint on Hugging Face + ModelScope, and a project page; the 14B causal-pretrain, 14B bidirectional, and 1.3B variants are on the TODO list [README §TODO].
The codebase is built on top of Wan2.2, and deployment guidance points to SGLang’s LingBot-World cookbook and NVIDIA’s flashdreams repo rather than in-repo serving code [README §Deployment, §Quick Start].
Inference is chunk-by-chunk with KV caching (generate.py), with tunable local_attn_size (default 18) and sink_size (default 6) parameters exposed on the CLI — matching the “attention sink + local window” long-context pattern used in Context Forcing: Consistent Autoregressive Video Generation with Long Context [README §Inference].
License is CC BY-NC-SA 4.0 (non-commercial, share-alike), a step down in permissiveness from v1’s Apache-2.0 release [README §License].
Live public demos are hosted by third parties — Reactor (international web) and LingGuang (domestic mobile) — with the “official setup” of the model reserved for WAIC 2026 [README §Try it now].
Reference sample invocation is 8-GPU (--nproc_per_node=8, --ulysses_size 8) with --dit_fsdp --t5_fsdp and 361 frames at 480×832 for a causal-fast rollout [README §Inference].

Method

The public release documents only the inference surface; the tech report referenced in the README (“Jul. 9, 2026: We release the technical report, inference code, and models”) has no working arxiv URL at filing time and the project page did not return content through the fetch. What can be reconstructed from the repo:

The base architecture inherits v1’s Wan2.2-derived DiT backbone. The v2 headline change is a causal pretraining paradigm replacing v1’s block-causal distillation-only causal path — v1 only got causality via the LingBot-World-Fast student, whereas v2 makes causal generation the training objective of the base model as well. The distilled causal-fast variant sits on top of this and reaches 720p / 60 fps. Inference is a chunkwise KV-cache autoregressive loop (generate.py) with explicit local-attention window and attention-sink tokens as CLI knobs — the same structural machinery (Attention Sink + Slow/Fast Memory + bounded RoPE) that Context Forcing: Consistent Autoregressive Video Generation with Long Context argued is what actually stabilizes long-horizon streaming.

Conditioning surface is different from v1: v1 exposed Plücker camera embeddings + multi-hot W/A/S/D keyboard vectors; the v2 README implies a text-driven action vocabulary with combat-style verbs (attacking, archery, spellcasting, shooting) alongside “richer text-driven events” — suggesting a shift from low-level control channels to language-mediated action semantics, though the mechanism is not documented.

The agentic harness is the most novel structural claim: two agents wrap the world model at inference time — a pilot agent that plans and executes character behaviour (i.e. synthesizes the action stream), and a director agent that generates novel environmental elements as the rollout progresses. Structurally this pattern echoes the “WFM-as-interruptible-tool inside a VLM orchestrator” pattern that VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation introduced for embodied manipulation, but applied to interactive video/gameplay generation rather than physical control.

Results

Real-time interactivity: 720p @ 60 fps with sub-second control latency, up from v1’s Fast variant at 480p / 16 fps — an order-of-magnitude improvement in visual throughput at a step up in resolution [README §Rapid Response Time].
Horizon: characterized as unbounded rather than v1’s “up to ~10 minutes”; no VBench-style benchmark numbers or specific horizon measurements are given in the README at filing time.
Live demos: two public web/mobile demos (Reactor / LingGuang) confirm the real-time claim is deployable in practice; the “full-capability” official demo is deferred to WAIC 2026 [README §Try it now].
Release footprint: 14B causal-fast checkpoint on Hugging Face + ModelScope + inference code at filing (July 9, 2026); causal-pretrain, bidirectional, and 1.3B variants pending [README §TODO].
Ablations and comparative benchmarks are not disclosed in the README — no VBench, no Genie-3 comparison, no side-by-side with v1 or with HY-World-1.5 / Yume-1.5, unlike v1 which reported specific VBench dynamic-degree numbers.

Why it’s interesting

For the wiki, LingBot-World 2.0 is the direct sequel to Advancing Open-source World Models (LingBot-World) and the newest datapoint on how the open Wan-based interactive-WFM recipe evolves — it moves from “block-causal distillation only” to “causal pretraining as the base objective”, and from 480p / 16 fps to 720p / 60 fps, positioning as a more credible open-source Genie-3 alternative. Compared to Project Genie: Experimenting with infinite, interactive worlds (closed flagship, ~20–24 fps at 720p) the throughput claim is now higher than the closed flagship’s public spec while remaining fully open-weight — the first filed datapoint where an open interactive-WFM release quotes better real-time numbers than DeepMind’s closed reference. The agentic harness (pilot + director agent) is also a structurally new pattern for this cluster: it’s the first LingBot entry to bundle an agent loop with the world model, and it sits closer to VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation‘s WFM-as-interruptible-tool framing than to the pure rollout WFMs of v1 or Genie 3. Two caveats worth flagging: (a) no tech-report URL resolves at filing, so the horizon and quality claims are unverified against benchmarks yet, and (b) the license drops from Apache-2.0 (v1) to CC BY-NC-SA 4.0 non-commercial — a meaningful narrowing of “open” that likely reflects productization pressure from the WAIC 2026 demo + third-party hosting deals.