Tencent open-sources Hy-Nomas interactive video world model that runs on a single RTX 4090
AIFlow_ML · x · 2026-10-25
Tencent Hunyuan released Hy-Nomas (Apache 2.0), a video world-generation model supporting free-form multi-turn interaction via text and camera trajectories, generating a persistent world segment by segment. Architecture: segment-autoregressive DiT + Qwen text encoder + camera conditioning + PixelUMM decoder, with a persistent KV cache carrying context across segments; distilled inference uses 8 denoising steps for 720×1280@24FPS. An INT8 DiT + INT8 KV cache + FP8 decoder + CPU offload configuration runs on a single RTX 4090. Ships with Hy-Nomas-Bench (280 test cases). AIFlowML flags the 'Model Assets' section for INT8/LoRA details and hints at "one catch".
More from Multimodal
- ComfyUI Krea2 Prompt Styler Node Revived in a Fork With a Massive Style Library — Winter_unmuted · 2026-10-27
- Single HTML file generates retro poster designs in-browser with zero LLM calls — Afinetheorem · 2026-10-27
- A Redditor Used Copilot to Imagine and Archive Their Own '80s/'90s Winter Streetwear Concepts — Valuable-Sky2134 · 2026-10-27
- Transkun open-source piano transcription model nails MIDI, sustain events included — alexanderchen · 2026-10-27
- fal adds FLUX 3 Image Klein LoRA training: 10 reference images, 3 LoRAs per request — OdinLovis · 2026-10-27
- Guy turns his girlfriend's selfie into a playable AI game ad about popping pimples — CurieuxExplorer · 2026-10-27