Hour‑Level World Model Unveiled: China’s LingBot‑World 2.0 Goes Open‑Source

Ant Alpha (蚂蚁灵波) announced on July 9 that LingBot‑World 2.0 is now open‑source. The model moves world‑model generation from the “minute‑level” to the “hour‑level”, becoming the first real‑time interactive world model that can run continuously for an hour or longer without visible quality decay. It solves the long‑time drift problem by redefining world simulation as a causal generation process, using a causal pre‑training paradigm and a self‑developed MoBA mechanism that predicts future frames from already observed scenes. Two versions are released: a 14 B parameter main model for high‑quality research and a 1.3 B distilled model that runs on consumer‑grade GPUs, both delivering stable 720p @ 60 fps output.

On July 7 OpenAI launched the GPT‑5.6 family, comprising three variants—Sol (flagship), Terra (balanced) and Luna (high‑speed). Pricing is Sol $5 / M input, $30 / M output; Terra $2.5 / M input, $15 / M output; Luna $1 / M input, $6 / M output. GPT‑5.6 Sol set a new high of 53.6 points on the Agent’s Last Exam, 13.1 points above Claude Fable 5, and achieved 80 points on the Artificial Analysis Coding Agent Index with less than half the output tokens of Fable 5, cutting runtime and cost by roughly one‑third. The series also introduces a parallel multi‑agent workflow that can run up to four agents simultaneously, improving scores on benchmarks such as BrowseComp, SEC‑Bench Pro and Terminal‑Bench 2.1. The accompanying productivity tool ChatGPT Work merges ChatGPT with Codex, enabling automatic creation of tables, slides, documents and web apps from context in Slack, Notion, Microsoft 365 or Google Drive, and supports scheduled tasks and a built‑in browser for web‑based research.

Tencent’s mixed‑expertise (MoE) model Hy3 was released with 29.5 B total parameters, 2.1 B active parameters and a 256 K context window. Compared with its predecessor, Hy3 shows notable gains in coding and agent capabilities, reduced hallucinations, and better product experience. Integrated into the AI assistant 元宝, Hy3 now lets users generate PPTs, Word documents, Excel sheets and PDFs directly in chat, with rapid iteration and context‑aware edits.

At the 2026 China Internet Conference, China Mobile showcased a full‑stack “digital‑intelligence” foundation that combines communication, compute and AI. Its MoMA model hub aggregates over 300 mainstream models with intelligent routing, while the intelligent‑agent sandbox, 磐基 PaaS and 磐维 database provide end‑to‑end data storage, compute scheduling and secure execution. The Nine‑Day MaaS platform unifies heterogeneous compute resources, delivering industry‑leading training and inference performance for 19 state‑owned enterprises. An anonymized data space preserves 90 % of raw data value while protecting privacy, serving thousands of government and corporate clients.

DeepSeek announced a V4 release with a peak‑hour pricing scheme that doubles API costs during weekday peak windows (9 am‑12 pm, 2 pm‑6 pm). The move reflects sustained demand and tightening compute supply, signaling a shift from pure price competition to cost‑efficiency and commercialization capabilities.

MiniMax introduced the M3 model (428 B total, 23 B active parameters) with a mixed‑pricing of $0.22 / M tokens—significantly cheaper than DeepSeek’s $0.35 / M after the hike. M3’s utilization exceeds 90 % and its revenue grew from $100 M ARR in Dec 2025 to $150 M in Feb 2026, with a projected $1 B ARR by year‑end, driven by a lean team where >80 % are R&D staff.

OpenAI also released GPT‑Live (versions 1 and 1 mini) on July 8, a full‑duplex voice model that can listen and speak simultaneously. Unlike earlier cascade or single‑model voice pipelines, GPT‑Live decouples interaction (continuous listening and short acknowledgments) from thinking (delegating complex queries to GPT‑5.5 or other back‑end models), enabling real‑time translation and smoother conversational flow.

In a recent podcast, Codex lead Andrew Ambrosino emphasized that the AI era’s most valuable skill is judgment—deciding what to build—since implementation costs have plummeted. He warned against over‑reliance on specific tooling and highlighted the need for designers to focus on outcomes rather than process.

Claude Code’s backstory was detailed: originating from an internal VS Code extension in 2021, evolving through reinforcement‑learning‑based function‑calling research, and eventually becoming a production‑ready tool after a 2024 green‑light. By early 2025 the Claude CLI (renamed Claude Code) was publicly released, and today it handles the majority of Anthropic’s coding workload.

Sergey Brin discussed Transformers at the DeepMind Build Day, noting that modern Transformers have diverged substantially from the 2017 design and now support multimodal tasks. He argued that while Transformers can still be a path toward AGI, true AGI will require embodied interaction and physical world understanding, making world‑model research critical.

A leak of Google’s Gemini 3.5 Pro (expected July 17) revealed a model re‑trained from a new foundation. Early tests show it excels at front‑end and visual code generation—producing pixel‑perfect UI, SVG and layout code that surpasses Anthropic’s Fable 5—but it lags behind in heavy reasoning, large‑scale software engineering and multi‑step agent tasks.

The Cannes AI Craft competition, introduced in June 2026, awarded two LingBot‑AI generated ads: “L'Ultimo Uomo Reale” (Silver in Classic Film, Bronze in Craft Film) and “Lorem Ipsum” (Bronze in Classic Film). These wins demonstrate that AI‑generated video now meets commercial‑grade standards for performance, consistency, and artistic direction, and LingBot‑AI reported Q1 2026 revenue of ¥6.5 B with over 100 M users worldwide.

Overall, the week highlights a rapid escalation in AI model capabilities—from hour‑level world simulation and real‑time voice translation to cost‑effective large‑scale models—while market dynamics shift toward pricing rationalization, multi‑agent productivity, and the commercialization of AI‑generated media.