Tencent Hunyuan open-sources Hy-Nomas, an interactive video world model steered by text and camera, runnable on one RTX 4090
Confirmed
The takeaway
A low-memory INT8 configuration runs on a single RTX 4090; the team also released Hy-Nomas-Bench (280 test cases).
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 26 October 2026
Your AI and this story
- GPT-6.2 Astra143 days after its cutoff
- Claude Opus 5.5113 days after its cutoff
- Gemini 4 Hydrogen204 days after its cutoff
- Grok 4.7143 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 113 days before it.
Key facts
- Repos: github.com/Tencent-Hunyuan/Hy-Nomas and huggingface.co/tencent/Hy-Nomas, both created Oct 21, 2026; a Diffusers port followed on Oct 24; Apache 2.0; ~206 GitHub stars on Oct 26
- Architecture (model card): segment-autoregressive diffusion transformer, Qwen text encoder, camera-conditioning modules and a PixelUMM video decoder (36 layers, 4,096 hidden); persistent KV cache between segments
- Output: 720×1280 at 24 fps; distilled inference with 8 denoising steps (4 in the Diffusers example)
- Interaction: multi-turn free-form instructions (add objects, change style or scene, describe actions) plus camera control
- Hardware: standard setup uses 8-way sequence parallelism (eight GPUs); a single-GPU configuration with INT8 DiT, FP8 decoder, bounded 18-frame history and CPU offload targets one GeForce RTX 4090
- Benchmark: Hy-Nomas-Bench, four suites and 280 test cases (multi-turn text interaction, camera navigation, interleaved text and camera control, spatial memory); web demo at hy-nomas.hy.tencent.com
What happened
Tencent released, without a press push we could find, an open interactive world model in the same line as Google’s Genie and Skywork’s Matrix-Game: users “walk” a generated world by giving text instructions and camera paths, and the model extends the video segment by segment while remembering what it showed earlier. Weights, code, a Diffusers port and a benchmark are public.
Why it matters
Open world models that run on a consumer GPU are rare. This one comes from a major lab under a permissive licence. Its quality relative to Genie-class systems is unmeasured outside Tencent’s own benchmark.
Sources
3 sources from 2 sites. Numbers match the chips in the text.
3 sources: 3 primary
Primary
- Hugging Face: tencent/Hy-Nomas model cardhuggingface.co, official
- GitHub: Tencent-Hunyuan/Hy-Nomasgithub.com, code
- Hugging Face: tencent/Hy-Nomas-Diffusers (Oct 24)huggingface.co, code
Changes
- Filed (found by a Hugging Face check of Chinese lab orgs; model card and repo read directly)