The other path for world models: NVIDIA puts Cosmos 3 post-trained checkpoints on a download page

On November 5 NVIDIA announced on the official Hugging Face blog four open-weight checkpoints of the Cosmos 3 family: Cosmos3-Nano-Transfer-Auto (multi-view driving data generation, up to 11 camera views with RGB and LiDAR generated jointly), Cosmos3-Nano-Sim-Auto (a distilled streaming variant of camera-plus-LiDAR observations, emitting small chunks in a few denoising steps), Cosmos3-Nano-Sim-Transfer (simulator controls plus text prompts to photorealistic video) and Cosmos3-Nano-Sim-Bimanual (robot actions to video, supporting YAM and AgiBot embodiments). All four are post-trained variants of Cosmos 3 Nano, 16B trainable parameters each, under the OpenMDW1.1 license. Our November 7 HF API check: the four repos were created August 20/31, weights updated October 30, 32.4–34.5 GB per repo — the November 5 blog post announces weights that had already been sitting in the repos.

The GA bases are open too: the Cosmos3-Super/Nano/Edge base repos carry open weights as well (created March 10, Edge July 1, updated September 16), with Nano at 137k downloads, Super 38k and Edge 64k at observation — these weights are already in real use, not an empty shelf. The runtime conditions are stated plainly: Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit; the vendor demo environment is one GB200, and the blog itself notes "this is not a minimum-memory specification". Across the aisle sits October 27's Genie 4: stronger emphasis on interaction (keyboard walking, text-driven scene edits, native audio), but no public weights, trusted testers first, then Project Genie inside the AI Ultra subscription.

• See the four post-trained checkpoints ↗
• See the Cosmos3 family collection ↗

On November 7 we read all 15 rows of the Physics-IQ Verified physical-video board's v2v · verified · all-listings view in a browser. The two catalogued Odyssey rows are unchanged: Odyssey-3 Pro BoN×8 66.10 first, Odyssey-3 BoN×8 64.43 third. What is new are the Cosmos rows: Cosmos 3 Super (Astra) BoN×8 65.47 second, Cosmos 3 Nano (Astra) BoN×8 61.83 sixth, and the same weights without prompt enhancement in single-run rows — Cosmos3 Super 50.80, Cosmos3 Nano 43.00. The "OPEN SOURCE" label now sits this high on this board: a downloadable model scoring just under the top spot on verified physics continuation, with the sixth-place Nano as the board's other open-weights row.

The configuration gap deserves its own reading: the highest Cosmos rows carry an (Astra) suffix, and the board's hover note says it is GPT-6-Astra prompt upsampling — an OpenAI model writing prompts for Cosmos. Strip the enhancement and sample once, and the same weights score 50.80 / 43.00. The distance from 43.00 to 61.83 lives in prompting and sampling configuration, not in the weights — half of an open-weight model's score comes from its harness. The same view also has two proprietary FLUX 3 [large] rows (Black Forest Labs) and two Microsoft Research Asia Physical Registry rows (marked NOT YET AVAILABLE — method public, model unreleased), a reminder that score tables can contain rows you cannot obtain.

Genie 4 still has no row: the gap recorded on October 28 stands in our November 7 read, and the view has no Google rows at all — the Veo / Gemini Omni Flash rows from the October 28 record are not in this 15-row layout. The row set has changed since October 28's 26 rows; the source-side change is recorded as-is. World models still lack an independent evaluation covering interactive capability; this physical-continuation board measures how environments change after actions, and Genie 4's audio and live scene editing sit outside its scope. When a fitting independent evaluation appears, this site will add it under the board's rules rather than pre-computing a rank.

• See the current Physics-IQ Verified board ↗
• Compare: Genie 4's board-gap section →

The technical report "Interactive World Simulation with Cosmos for Physical AI" (PDF dated November 4) supplies the vendor's own numbers. Robot benchmarks use 232 AgiBot-World windows (186 source episodes) plus 100 AgiBot-GEAR windows (30 source episodes): the final SF-DMD student of Sim-Bimanual (4 denoising steps, a K/V window of 64 with 8 sink frames) beats all listed DreamDojo teachers on five metrics across both benchmarks — against DreamDojo 14B, PSNR rises 2.06 dB (World) and 1.90 dB (GEAR) while LPIPS drops about 3.7% and 6.2%. A 16B-trainable post-train outperforming 14B references on vendor-chosen benchmarks is the report's most direct selling line.

Sim-Transfer raises all four spatial-control metrics against the Cosmos 3 GA reference: segmentation IoU 0.6433→0.7378, blur SSIM 0.8559→0.8867, Edge F1 0.2140→0.3407 (+59.2%) and depth SI-RMSE 0.8423→0.6194 (−26.5%); VLM Quality moves 4.9296→4.9508. But the costs sit on the same page: DOVER and Prompt drop 1.8% and 2.2%, quality drift and Sampson geometric error rise, and across distillation stages the bidirectional teacher stays stronger than the final causal student on all nine metrics — streaming-causal realtime is bought with metrics.

Every number here is NVIDIA's own reporting, not an independent reading. The report's own boundaries deserve keeping whole: the public benchmarks evaluate action-conditioned observation prediction, "work remains to measure where the models faithfully support policy learning and evaluation"; closed-loop interaction needs a simulation runtime and a WSM renderer; and "a successful short run verifies execution, not convergence". This site downloaded no weights, ran no model and reproduced nothing — read 65.47 as the independent physics-board reading and the report figures as vendor-reported, and do not mix the two.

• Read the technical report ↗

Set Genie 4 beside Cosmos 3 and the two release paths for world models become clear. Google takes the subscription side: the model exists only inside its interface, no weights, no independent board row, SynthID watermark, with capability staked on interaction — walk in with a keyboard, edit the world with text, native audio, about two minutes of scene coherence. NVIDIA takes the download side: 32GB-class checkpoints under OpenMDW1.1, aimed at autonomous-driving and robotics simulation research, requiring Linux, an NVIDIA GPU and Docker, with a GB200 as the vendor demo environment; capability is staked on data generation and policy evaluation, not on interactive experience. The audiences differ too: those who want to step into a world join Genie's waitlist; those who want to fit a world into a training loop go to Cosmos's download page.

Three actionable points for readers: first, to know where open weights stand on independent physics evaluation, Cosmos 3 Super's 65.47 is the highest "OPEN SOURCE" row on the board today (the other is Nano at 61.83); second, those with a GPU and their own data can download the checkpoints — but the vendor itself says a successful short run is not convergence, and post-training needs your data and your verification; third, for the step-into-a-world experience, Genie 4's entry remains the trusted-tester waitlist and the AI Ultra subscription. Of our boards only the physical-continuation one gives world-model readings; AA and Pi target language and coding models, and the calibers do not convert.

• Next: the full Genie 4 article →
• Next: how to read physical-video scores →