On November 5, Nvidia released four new open-weight checkpoints from the Cosmos 3 family on Hugging Face. All are post-trained from Cosmos 3 Nano, each with 16 billion trainable parameters, and licensed under OpenMDW 1.1—meaningcommercial use is permitted.
Another piece of news on the same day: funding for world models is drying up. According to Qbit News’ November 8 report, first-tier investors’ attitude toward world models has shifted from scrambling for projects in the first half of the year to struggling to issue term sheets (TS) in the second half; only a handful of projects have attracted commitments from top-tier funds, and there were just five funding events related to physical AI and world models in October.
On one side, Nvidia is open-sourcing its most expensive technology for free; on the other, capital is flowing out. Viewed together, these two developments are more intriguing than either one alone.
What each of the four checkpoints does
This release is not a general-purpose world model, but rather four specialized variants tailored for physical AI—each addressing a specific stage:
- Cosmos3-Nano-Transfer-AutoBidirectional model dedicated to offline synthetic data generation for autonomous driving. Generates up to 11 synchronized camera views (including fisheye) and supports joint RGB and LiDAR generation; scene structure and traffic motion are controlled via the World Scenario Map (WSM), while appearance is controlled via text prompts. For autonomous driving teams, this is a tool for “creating long-tail scenarios without real-world vehicle deployment.”
- Cosmos3-Nano-Sim-AutoCausal distillation streaming version of Transfer-Auto. Generates only short segments of camera video and LiDAR per step, with minimal denoising steps; advances forward using rolling historical context plus new driving control inputs—ideal for long-horizon closed-loop simulation.
- Cosmos3-Nano-Sim-TransferTranslates control signals from physics simulators into RGB images. Physics and ground-truth state remain inside the simulator; the model handles only rendering visuals interpretable by the policy.
- Cosmos3-Nano-Sim-BimanualAction-to-vision prediction for dual-arm robots. Supports both YAM and AgiBot embodiments, and can be driven either by policy or manually via joystick.
What’s truly valuable technically
Not “generating prettier video,” butStreaming generationTraditional video models produce clips in bulk; world models require “I move, and the world changes accordingly”—each observation influences the next action, and each action in turn alters subsequent observations. Nvidia’s Sim-Auto compresses this loop into just a few denoising steps, enabling real-time closed-loop operation.
Another key is multi-sensor consistency. The core pain point in autonomous driving has never been “ugly visuals,” but misalignment between what cameras see and what LiDAR measures. Transfer-Auto supports joint RGB and LiDAR generation, enforcing geometric and appearance self-consistency—a capability pure video models lack and the hardest-to-replace aspect of these checkpoints.
Nvidia also released post-training recipes and an open agent skill (directly executable by coding agents for inference). For the “multimodal world model” we previously discussed in NVIDIA Cosmos 3 Deep Reviewthis is a deconstruction of the foundation—handing users full control over the modifiable components.
Who benefits most
First are autonomous driving and robotics teams—especially small and mid-sized ones. Previously, generating synthetic data meant either purchasing closed-source APIs on a per-use basis or training models in-house—both costly, with the latter demanding prohibitive compute. Now, 16-billion-parameter weights are downloadable outright; adaptation cost shifts from “training from scratch” to “fine-tuning.”
The second category comprises companies developing embodied intelligence simulation. Sim-Bimanual natively supports both the YAM and AgiBot embodiments, indicating that Nvidia built it specifically for hardware already shipped within the ecosystem. Domestic dual-arm robot teams can use it to predict action outcomes and reduce physical trial-and-error.
The third category consists of research institutions. Four checkpoints plus technical reports collectively lower the barrier to reproducibility in the world model field. For comparison: Google Genie 4 This contrasts sharply with approaches that rely solely on product demos while keeping weights closed—two fundamentally different openness strategies.
On the other hand: capital is exiting the world model space.
Several figures in QuantumBit’s report highlight the issue clearly: in the first half of the year, the term “world model” itself carried a valuation premium—many companies secured three consecutive funding rounds based purely on demos, with each round priced higher than the last; by the second half, that premium “shrank rapidly,” and investors now demand technical evidence, research depth, and tangible application progress—demo videos alone no longer attract interest.
A harder benchmark comes from the secondary market. Unitree Technology’s stock surged to RMB 1,100 on its first trading day, valuing the company at approximately RMB 444.9 billion; by October 15, it had fallen to RMB 404. Such valuation corrections are typical even for leading robotics firms—so how should investors value model startups that haven’t yet demonstrated scalable applications? They already have their own internal benchmarks.
This contrasts with AMD’s $8.2 billion acquisition of World Labs —a deal announced at the end of September, when the industry was still paying a premium for “spatial intelligence.” In less than two months, sentiment shifted entirely.
My assessment:
Open weights and capital retreat are two sides of the same coin. When top-tier foundational models are freely released by major tech firms, a startup’s claim of “we trained a world model” loses intrinsic value. Remaining value resides only in two places: first, vertical-domain data and closed-loop systems (e.g., real-world operational data from a specific class of robots); second, packaging this technology into shippable, production-ready products.
For Nvidia, this is a guaranteed win: open-sourcing models doesn’t diminish GPU sales—running those models still requires buying cards. For founders, the moat shifts unavoidably from “model” to “scenario”—which isn’t necessarily negative, but does close the window for narrative-driven fundraising.
Frequently Asked Questions (FAQ)
Can Cosmos 3 Nano weights be used commercially?
Yes. All four checkpoints are licensed under OpenMDW 1.1, which explicitly permits both commercial and non-commercial use. Specific terms are governed by the license text on the Hugging Face model page. Before commercial deployment, verify compliance requirements for your use case—especially when real vehicle or robot data is involved.
What hardware configuration is required to run Cosmos 3 Nano?
With 16 billion parameters, local inference demands a high-memory GPU (A100/H100-class recommended for reliability); streaming simulation imposes even stricter memory and bandwidth requirements. Consumer-grade GPUs cannot realistically run the full model—cloud GPU deployment is advised.
How do world models differ from video generation models?
Video models aim to “generate an aesthetically pleasing video,” whereas world models aim to “predict the consequences of actions” and sustain continuous closed-loop operation. The former concludes after one-time generation; the latter feeds each step’s output back into the next, demanding entirely different levels of consistency and speed.
Is there still opportunity in entering the world model space today?
Opportunities for general-purpose world models are now extremely slim—the foundational models have been open-sourced and offered free by major players. Real opportunities lie in vertical domains: operational data from specific robot classes, industry-specific simulation environments, and engineering the model into deliverable, production-grade products. Today’s funding market values only technical evidence and real-world deployment progress—not conceptual appeal.
Summarize
英伟达这次放出的四个 checkpoint，把世界模型从”只有大厂能训”推进到了”团队可以微调”。同一周里资本却在撤，说明行业正在从”讲概念”切换到”拼落地”。对做自动驾驶和具身智能的团队，这是难得的时间窗口——工具变便宜了，热钱变少了，正好是认真做事的时候。
想对比更多开源模型的规格和价格？查看我们的 AI Model Library and Four actionable countermeasuresor continue reading:NVIDIA Nemotron 3.5 Super VL Review · GPT-6 Astra 操控宇树 G1 · 小米开源 HachimoDock.