The World-Model Rush: LeCun's $1B Bet Meets Tencent's Open Weights

The "world model" idea stopped being a research footnote this week. On one end, Yann LeCun raised over $1B for AMI Labs to build models that understand the physical world, not just text. On the other, Tencent's Hy-Nomas text-to-video world model hit the top of Hugging Face trending, with a 9B Diffusers port and a ComfyUI wrapper appearing within days.

One thesis, two very different bets

Both pushes target the same gap: today's frontier models are impressive at language but weak at predicting how physical systems behave. LeCun's lab is an ex-Meta frontier bet, funded to go after that problem from scratch. Tencent, meanwhile, is shipping open weights that researchers and local-runners can grab immediately.

Nvidia is in the same neighborhood: its Cosmos3-Nano-Policy-DROID world-action model for robotics policies has drawn thousands of likes on Hugging Face. Meanwhile, Liquid AI's LFM3 - 2B multimodal models with 256K context that run in about 1GB at 4-bit - suggests the edge of these capabilities is moving onto devices, not just datacenter GPUs.

The money behind the models

The world-model race is landing in the middle of a capex supercycle. Microsoft reported $92.3B in Q1 revenue with Azure up 48%, and guided FY27 capex to $175B - a 10-Q that also disclosed $10B in OpenAI revenue, 11% of the quarter. Alphabet followed with Google Cloud up 103% to $30.7B and 2027 capex set to top $300B.

That spending is producing frontier output too: Google shipped Gemini 4 Carbon, which tops the AA Intelligence Index at 59, with a cyber-defense focus aimed at enterprise security.

The local-inference angle

For anyone running models at home, the HF trending list is unusually actionable: RecursiveSI/flint (40B, 14k likes) at Q4–Q6 fits comfortably in 128GB of RAM, and Nanbeige4.3-33B sits right in the 27–33B class that runs easily on a 128GB box at Q8. KRAFTON's Raon-1-mini-preview shipped with an official GGUF port trending on day one - 29B, Ollama-ready out of the box.

Meanwhile, the broader agent-tooling scene is consolidating around a different problem: morluto/rea gained 16.5k stars in a single day for agent-driven reverse engineering down to native binaries, while guardrail-style repos like hallmark and impeccable signal that the frontier has shifted from "build an agent" to "make agent output less slop."

Signal to watch: Moonshot's Kimi K3.1 Pro open-weights release on November 4 - it lands right in the 33B class where local boxes are competitive, and it'll be the first real open test of whether world-model-era training recipes show up in practical weights.