Cosmos 3 Nano vs Genie 4: Open Weights vs Closed Preview

NVIDIA published four new Cosmos 3 Nano checkpoints on 5 November 2026, each with 16 billion trainable parameters, as open weights for driving and robot simulation. Google's rival, Genie 4, arrived eleven days earlier, on 27 October, as a closed research preview you cannot download yet. A world model is an AI system that predicts how an environment will change and how your actions will affect it, by generating the next moment of a scene rather than the next word. These two are built for opposite jobs: one is a tool you can run, the other is a demo you can watch.

What NVIDIA shipped on 5 November

The 5 November release is four checkpoints, not one model: post-trained versions of Cosmos 3 Nano, each with 16 billion trainable parameters, under the OpenMDW1.1 licence. NVIDIA also published the post-training recipes and a technical report. Cosmos 3 itself launched on 31 May 2026, so only these four are new.

Each checkpoint does one job. Cosmos3-Nano-Transfer-Auto is a bidirectional model for offline synthetic driving data: synchronised video across up to 11 camera views, including fisheyes, with joint RGB and LiDAR generation. World Scenario Map controls set scene structure and traffic motion; text prompts guide appearance. Cosmos3-Nano-Sim-Auto is the causally distilled streaming variant, generating successive small chunks of camera video and LiDAR in a few denoising steps, so a simulation runtime can close the loop over long rollouts.

The other two target simulation and robotics. Cosmos3-Nano-Sim-Transfer turns a physics simulator's spatial controls — depth, segmentation, edges, blur — plus a text prompt into photorealistic video a policy can consume. Cosmos3-Nano-Sim-Bimanual simulates how an environment evolves in response to a two-armed robot's actions, for YAM and AgiBot embodiments, controlled by a policy or joysticks.

Under the hood, Cosmos 3 uses a Mixture-of-Transformers backbone joining text and visual networks, trained in four stages ending with self-forcing on generated history. The setup requires Linux, an NVIDIA GPU, Docker, and NVIDIA Container Toolkit. It was tested on one GB200 — not a minimum-memory specification, NVIDIA states.

Illustration created for INORYVA

What Google shipped on 27 October

Google DeepMind Genie 4 is a different kind of product: a real-time interactive world model. You type a description of a place, and it renders that place live, with sound, while you steer it with a keyboard and more text. Genie 4 generates audio, music, ambient noises, and interactive dialogue with the world — DeepMind's first real-time world model to generate audio natively. Every output carries SynthID, Google's imperceptible watermark.

Genie 4 is built on the Gemini family. Gemini Omni anchors simulations in real-world appearance, while Gemini 3.8 Flash serves as a guiding agent in the loop, observing the scene and adapting its plan as the world shifts. DeepMind also showed Gemini 4 Argon with Genie 4 simulating Apptronik's Apollo humanoid playing checkers — an early step toward hybrid world models.

The access story is the opposite of NVIDIA's. Genie 4 rolled out as a research preview for trusted testers, coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio. Everyone else joins a waitlist: no download, no public API, no price. Reports put consistent scenes at about two minutes, roughly double Genie 3, but that figure comes from TechCrunch and others, not DeepMind's own announcement, which says only that Genie 4 holds coherent environments over short sessions. DeepMind lists the limits openly: drift, complex physics, audio accuracy, legible text, and long-term consistency are still open research.

Head to head: the table that matters

Detail	Cosmos 3 Nano checkpoints	Genie 4
Announced	5 Nov 2026 (four checkpoints)	27 Oct 2026
Access today	Open weights, downloadable	Research preview only
Licence	OpenMDW1.1	Closed, no licence published
Size	16B trainable parameters each	Not published
Native audio	Video and LiDAR only	Yes — music, ambient sound, dialogue
Main use	Driving and robot simulation	Interactive worlds and games
Hardware	Linux, NVIDIA GPU, Docker	Not published
Price	Free download	Not published
Benchmarks	None published with the checkpoints	None published
Known limits	Policy fidelity unmeasured	Drift, physics, audio, text

The table makes the real difference obvious. Cosmos 3 Nano is a tool for builders: download it, fine-tune it on your own data, run it in your own simulation, and measure whether it works. Genie 4 is a tool for experiencing: watch DeepMind's demonstrations, join a waitlist, and imagine what you would build — but you cannot build with it today. NVIDIA ships infrastructure; Google previews an experience.

One discrepancy is worth flagging. NVIDIA's blog says 16 billion trainable parameters per checkpoint, while secondary coverage notes the Hugging Face model cards list 15 billion. This article uses the blog's figure, as the primary source, and notes the difference so readers checking the cards are not confused.

Illustration created for INORYVA

Different jobs, different buyers

For an autonomous-driving team, the choice is not close. Transfer-Auto generates multi-view driving data, and Sim-Auto streams synchronised camera and LiDAR observations for closed-loop simulation. The honest limit is NVIDIA's own: further work is needed to measure where the models faithfully support policy learning, and the sim-to-real gap is still unsolved. Validate on held-out real episodes before trusting the simulation.

For robotics, Sim-Bimanual and Sim-Transfer are the interesting pair. Sim-Transfer lets a physics simulator keep ground truth while the model renders what a policy would see. Sim-Bimanual predicts the next observation directly from the robot's actions — closer to a true world model, but riskier, since errors become the next input. Whether it works for real robot training is still open.

For a game developer, filmmaker, or teacher, Genie 4 is more exciting — and less usable. DeepMind's trusted testers already use it as a rapid draft mode for pre-visualisation, and photojournalists are testing whether one archival photograph can become a grounded simulation of a historical space. Those are real uses, but they belong to a small group today. Our flagship AI model comparison and Gemini 4 Carbon cover the same season, but neither is a world model.

The open-weight question also splits the two. Mistral Large 4 showed the same pattern in language models: open weights let anyone inspect, adapt, and deploy, while closed previews let the maker control the experience. Cosmos follows the open path, with a commercial-use licence and post-training recipes. Genie follows the closed path, with SynthID on every output. Both are defensible; they serve different users.

What neither model proves yet

Neither release comes with a benchmark to rank them, and this article will not invent one. NVIDIA published no scores, and DeepMind published no benchmark, resolution, or frame rate. One score would mislead anyway: Cosmos is judged on real-hardware results, Genie on whether people find its worlds convincing and steerable.

The shared limit is time. Cosmos rollouts depend on generated history, and errors compound over long sessions, which is why NVIDIA added self-forcing. Genie sessions drift too. A world model that holds together for two minutes is a demo; one that holds together for hours is infrastructure.

Illustration created for INORYVA

Key Takeaways

• Four Cosmos 3 Nano checkpoints, 5 Nov 2026 — Transfer-Auto, Sim-Auto, Sim-Transfer, Sim-Bimanual — each 16B trainable parameters, OpenMDW1.1.
• Google's Genie 4 (27 Oct 2026) is a closed research preview for trusted testers, with native audio and Gemini in the loop; no public download or price.
• Cosmos is built for driving data and robot simulation, usable today; Genie is for interactive worlds almost nobody can use yet.
• Neither vendor published benchmarks, prices, or full specs — no honest single score ranks one above the other.
• Both models drift over long sessions; measuring real-world reliability is unfinished work.

Conclusion

Not a race with one winner. NVIDIA gave builders downloadable weights, four checkpoints, and adaptation recipes, while admitting policy-learning fidelity still needs measuring. Google gave a glimpse of interactive worlds with sound, dialogue, and reasoning in the loop, while admitting drift, physics, and long sessions are still unsolved. To train a robot or driving policy this week, use Cosmos. To see where interactive worlds are heading, watch Genie 4.

Sources

• NVIDIA/Hugging Face blog, "New NVIDIA Cosmos 3 Open Models for Interactive World Simulation", 5 Nov 2026 — https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation (read in full, 7 Nov 2026)
• Google DeepMind blog, "Genie 4: A new generation of world simulation", 27 Oct 2026 — https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ (read in full, 7 Nov 2026)
• TechCrunch coverage of Genie 4, 27 Oct 2026, via secondary reporting (two-minute figure — reported, not DeepMind's text)