LingBot-World 2.0¶

Development line: project:lingbot-world-2 · thread lingbot-world-2-development
Last event: 2026-07-09 · 1 dated since 2026-07-09 · Researched: 2026-09-05 · confidence: high

What it is¶

LingBot-World 2.0, also called LingBot-World-Infinity, generates navigable image-to-video worlds for researchers and developers. It supports unbounded streaming, action conditioning, and a pilot or director agent harness. The public local checkpoint is robbyant/lingbot-world-v2-14b-causal-fast (14B). The repository documents 480p multi-GPU inference, while Reactor hosts reactor/lingbot-world-2 at 1664×960 and 48 fps. We use Reactor for prototypes and the 14B repository release for non-commercial research, not for self-hosted real-time deployment.

Development line¶

• 2026-07-09 — LingBot-World 2.0 public development resources were linked. On 2026-07-09, a record linked LingBot-World 2.0 to its project page, GitHub repository, Hugging Face collection, and Reactor page. This marks a public milestone, though the source does not detail release contents, capabilities, or operational status.

What changed¶

• 2026-07-08 — The technical report reached arXiv, describing the 14B model and a planned 1.3B counterpart.
• 2026-07-09 — LingBot-World 2.0 published its report, code, and models. The working open checkpoint was lingbot-world-v2-14b-causal-fast. The 14B causal-pretrain, 14B bidirectional, and 1.3B variants remained listed as TODO. The release added causal long-horizon generation, a distilled real-time variant, action and text controls, multi-player support, and a pilot/director harness.
• 2026-07-10 — The repository renamed a causal class for code maintenance, with no new capability.
• 2026-07-14 — The repository revised a video prompt description in documentation, without a new model release.

How to use this¶

As of 2026-07-09, treat LingBot-World 2.0 as its own line. Review the project page, repository, model collection, and Reactor docs before adopting it.

1. For a hosted prototype, scaffold a Reactor application with npx create-reactor-app my-lingbot-world-2-app --model=lingbot-world-2. — https://docs.reactor.inc/model-api-reference/lingbot-world-2/overview
2. Open a session with model name reactor/lingbot-world-2, upload a reference image, set a text prompt, then send start. Generation fails until both image and prompt are present. — https://docs.reactor.inc/model-api-reference/lingbot-world-2/overview
3. Steer the running world with longitudinal and lateral movement, look controls, or set_camera_pose. Update conditions mid-stream with set_prompt. — https://docs.reactor.inc/model-api-reference/lingbot-world-2/overview
4. For local research inference, clone the repository, install requirements with Torch 2.4+ and FlashAttention, download robbyant/lingbot-world-v2-14b-causal-fast, then run the provided multi-GPU causal-fast command or run_fast.sh. — https://github.com/robbyant/lingbot-world-v2

Best practices¶

• Separate prompts into base, camera, movement, event, and vertical-control layers. Keep each layer on one control axis so prompts do not fight active motion. — https://docs.reactor.inc/model-api-reference/lingbot-world-2/prompt-guide
• Keep total prompt length under roughly 2,000 characters. Trim event clauses before base identity or control rules. — https://docs.reactor.inc/model-api-reference/lingbot-world-2/prompt-guide
• Treat movement as persistent state and send an explicit idle command. Short taps can miss the chunk boundary. — https://docs.reactor.inc/model-api-reference/lingbot-world-2/overview
• Do not expect official deployment code for the open model. The repository withholds it, and the license is CC BY-NC-SA 4.0 for non-commercial use. — https://github.com/robbyant/lingbot-world-v2

Superseded by this¶

• 2026-07-09 — LingBot-World 1.0 notes are obsolete on Reactor. World 2 replaces them with two-axis navigation, camera poses, and mid-stream prompt steering.
• 2026-07-09 — Expecting every announced 14B and 1.3B model to download is obsolete. Only the 14B causal-fast checkpoint is public; the rest remain TODO.

Still unknown¶

• The July 9 post claims the real-time model runs 720p at 60 fps, while Reactor specifies 1664×960 at 48 fps. The sources provide no direct comparison.
• The repository withholds deployment code. Hardware requirements, serving stack, and latency for official real-time serving remain undocumented.
• The arXiv paper describes a 1.3B counterpart, but the July 9 repository TODO list provides no downloadable 1.3B weights or code.

Sources¶

Agent brief¶

• Subject: project:lingbot-world-2, thread lingbot-world-2-development, 1 dated events 2026-07-09 → 2026-07-09.
• Practical note: As of 2026-07-09, treat LingBot-World 2.0 as a distinct line. Review the project, code, collection, and product docs before adoption.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.