Tencent's Hy-Nomas world model adds image init and real-time streaming on B300
Scobleizer · x · 2026-11-11
Two weeks after release, Tencent's open-source interactive world model Hy-Nomas gets an update: reference-image initialization and real-time streaming on B300, with a demo at SGLang Summit.
Per the Hugging Face model card, Hy-Nomas is a video world generation model supporting free-form multi-turn interaction:
- Segment-wise autoregressive DiT: pairs a Qwen text encoder, camera-conditioning modules and a PixelUMM decoder, with a persistent KV cache carrying context across segments
- Free-form interaction: introduce objects, change styles/scenes, describe actions, or control camera trajectories during generation
- Distilled inference: 8 denoising steps per segment, outputting 720×1280 at 24 FPS
- Flexible deployment: sequence-parallel config or single RTX 4090 (INT8 DiT weights, INT8 KV cache, FP8 decoder, CPU offload)
Released under Apache-2.0 with a web demo and the Hy-Nomas-Bench benchmark.
More from Multimodal
- Dev builds a full trailer with Fable 5.2 in 10 minutes, prompt to final cut — FinanceYF5 · 2026-11-11
- Fable 5.2 use cases roll in: code-based painting, 4-min videos and a 10-min trailer — FinanceYF5 · 2026-11-11
- Nuance launches Nuplex model claiming step-change realism for AI personas, with safety plan — Scobleizer · 2026-11-11
- Fable 5.2 + Vork Renders 1080p Anime in 20 Seconds — All Code, No Video Model — ravisparikh · 2026-11-11
- Fable 5.2 Turns Scripts Into Animated Films, Shifting Input to Intent — FinanceYF5 · 2026-11-11
- Five Wild Fable 5.2 Use Cases, From 3D Art to 25,000 Agents Building Worlds — FinanceYF5 · 2026-11-11