Hy-Nomas: Tencent's open video world model you steer on a 4090
8B weights, Apache 2.0 and a live demo ship today — the open, runs-on-your-desk answer to Genie. Plus: Codex adds a native Google Vertex AI provider.
Copy markdown
Steer the video as it renders
Tencent Hunyuan open-sourced Hy-Nomas, an ~8B diffusion video world model you drive turn by turn with free-form text and camera moves. It builds each scene in segments and keeps a “lifelong KV cache,” so objects, lighting and geometry persist as you change things mid-shot — the open, Apache-2.0 answer to closed world models like Genie. Weights, code and a live demo are out now.
And it runs on a single RTX 4090
The repo ships a base DiT for LoRA fine-tuning, a distilled DiT for fast 4/6/8-step inference, and an INT8 build that runs on one RTX 4090 via KV cache, a sliding history window and CPU offload (Linux, Python 3.12+, CUDA 13). Two LoRAs are bundled — a cinematic look and a dual-arm robotics adapter trained on RealSource-World — so you can generate steerable worlds without a datacenter.
The catch: no technical report yet
Hy-Nomas shipped weights before numbers: the technical report is still marked “Coming Soon,” and the only public quality signal is the team’s own human evaluation — no head-to-head benchmarks against Genie, Wan or Veo. Great for prototyping and demos, but test it on your own prompts before building a product on top.
Codex 0.165.0 gets native Google Vertex AI
OpenAI’s Codex CLI 0.165.0 adds a built-in Vertex AI provider: set model_provider="gcp" to run Codex straight through Google Cloud — against your GCP commits, no proxy. The release also drops GPT-6.1 Terra into the model catalog, lets you change reasoning effort mid-conversation by default, and turns on memory v2 for fresh profiles.