HiDream-O1-World tops an interactive world-model benchmark

Share
HiDream-O1-World tops an interactive world-model benchmark

World models keep crossing out of the research demo into something you can actually steer — and the newest one to break the frontier comes from a Chinese lab better known for image generation. ZhiXiang WeiLai's HiDream-O1-World is an interactive, native all-modal world model, and its first trip onto a public benchmark just put it at the top.

HiDream-O1-World, the interactive world model from Chinese lab ZhiXiang WeiLai (the HiDream image-model maker), topped the Navi ranking of WBench — the new interactive-video world-model benchmark — with an average score of 80.9. The model generates a coherent, walkable world you can roam from first or third person, edit in real time, and build from text, an image, or direct interaction — a real-time sibling to what we saw from HelixWorld this morning (HelixWorld adds real-time sound to interactive AI worlds). What lifts it above the pack, per the lab and the benchmark numbers, is consistency: it scored 88.0 on spatial consistency and 73.3 on physics, the top mark on that axis. ZhiXiang WeiLai credits a native all-modal "UiT" backbone combined with memory of scene geometry and online test-time adaptation, so objects don't drift, vanish, or defy gravity as the camera swings — the classic failure mode that has made interactive worlds feel fake.

The headline number matters less than what it signals: WBench, built by Meituan's LongCat team with Fudan University, is the first real yardstick for interactive world models, spanning 289 multi-turn cases and 1,058 interaction rounds. HiDream-O1-World being the first to sit on top of it gives the field a reference point it didn't have. The underlying paper was also accepted to ECCV 2026. ZhiXiang WeiLai pitches the model at interactive film and games, embodied-AI simulation, and 3D scene production for creators — all natural next stops, but the honest read is that consistent, physics-respecting, long-horizon worlds are still the hard part, and topping one benchmark doesn't settle it.


ByteDance's Seedance 2.5 video model now generates native 1080P, and the API is open for it. The Volcengine-hosted model, pitched at "cinematic long-form" clips, had topped out at 480P and 720P; the new tier adds native 10-bit color output, sharper edges and fabric/hair detail, and more natural skin and lighting meant to cut the "AI look" from generated footage. It's a resolution and fidelity step rather than a new architecture, but it pushes long-form generation toward a place where output can actually pass through professional grading and post-production. The 1080P image-to-video API is live with a launch discount to roughly 72 percent of list price through September 17, which is ByteDance's way of seeding the higher tier into real workflows while the market for cinematic video models gets crowded.

What to watch: whether ZhiXiang WeiLai follows with open weights for HiDream-O1-World, and how quickly Meituan's WBench becomes the default leaderboard for interactive worlds.

Interactive world models can now hold a consistent, explorable scene — but do they need physics that's real before they're useful, or is believability enough? Write to us via our contact page.

Source: Leiphone — HiDream-O1-World

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Ornith 1.5 open models match Claude Opus 4.8 on coding benchmarks

Ornith 1.5 open models match Claude Opus 4.8 on coding benchmarks

An open-source model family just posted numbers that rival one of the most capable closed models in the world — and it's available under an MIT license. Ornith 1.5, released today by the Ornith AI team, is a three-model family spanning 9 billion dense parameters up to a 397 billion mixture-of-experts flagship. The big headline: the 397B MoE scores 86.1 on Terminal-Bench 2.1 (Terminus-2) and 86.0 on SWE-bench Verified, placing it on par with Claude Opus 4.8's 85.0 and 85.8 respectively.

Apodex launches TRACES, a benchmark that evaluates AI on real scientific discovery

Apodex launches TRACES, a benchmark that evaluates AI on real scientific discovery

Most AI benchmarks test whether a model can recall the right answer. TRACES, introduced today by Apodex, asks a harder question: can an AI system actually investigate a problem it hasn't seen before, form hypotheses, use tools, and arrive at a verifiable solution? The benchmark comes with a technical report on arxiv and a press release backed by multiple outlets, giving it more credibility than the usual "we made a leaderboard" announcement.

Detecting benchmark contamination that survives RL post-training

Detecting benchmark contamination that survives RL post-training

Every leaderboard number carries a quiet question: did the model learn the material, or memorize the exam? A preprint posted to arXiv this week gives auditors a sharper way to tell the difference — one built for the exact regime where today's detectors go blind. A new protocol proposes detecting benchmark contamination from a model's internal activations rather than its outputs — and it is designed to keep working after RL post-training, the stage that erases the signals existing detectors…

DeepSeek's V4-Flash-Vision rivals Opus 4.8 on agent tests

DeepSeek's V4-Flash-Vision rivals Opus 4.8 on agent tests

DeepSeek keeps moving fast on its low-cost track, and its latest move pairs image understanding with the company's signature price. Today it unveiled V4-Flash-Vision Exp, an experimental multimodal model that adds vision to its V4 Flash series, and DeepSeek says it stacks up against Anthropic's flagship on visual-agent benchmarks. DeepSeek's V4-Flash-Vision Exp nearly matches Opus 4.8 on multimodal agent tests. The new model, live on DeepSeek's API platform as deepseek-v4-flash-vision-exp…

Xiaomi's MiLM Plus ships PROVE, a benchmark for AI object removal

Xiaomi's MiLM Plus ships PROVE, a benchmark for AI object removal

The AI-editing arms race just got a proper measuring stick: Xiaomi's research team shipped an evaluation framework that says the metrics everyone uses to judge object-removal models are out of step with human perception. Meanwhile, the open-source fine-tuning favorite Unsloth put local model training on the desktop. Xiaomi's MiLM Plus research group released PROVE, an open evaluation framework for object removal in images and video that its authors say fixes what standard metrics get wrong.

vivo's 3B on-device model tops phone LLM chart, nears cloud scores

vivo's 3B on-device model tops phone LLM chart, nears cloud scores

A SuperCLUE benchmark released Friday found that a small model living entirely on a phone is competitive with the cloud. vivo's BlueLM 3.5 Nano 3B took the top spot on the group's on-device leaderboard with 89.86 points, edging past the field while scoring within a small margin of flagship cloud systems it wasn't even competing against. vivo's on-device BlueLM 3B topped SuperCLUE's phone-LLM chart — and nearly matched the big cloud models.