The 16B-trainable base of the Cosmos 3 family: 137k downloads on the GA repo at observation; four open-weight post-trained checkpoints (Transfer-Auto/Sim-Auto/Sim-Transfer/Sim-Bimanual) target 11-view + LiDAR driving simulation and bimanual robot actions to video, OpenMDW1.1, 32.4–34.5 GB per repo, vendor-reported wins over 14B teachers. Physics rows 61.83 (Astra + BoN×8) and single-run 43.00; AA/Pi do not apply.
The same verified video-to-video score, with Custom prompts in our recorded rows. The BoN×8 Cosmos rows carry (Astra) — the board notes GPT-6-Astra prompt upsampling; the unenhanced single-run rows score lower (Super 50.80 / Nano 43.00), and that configuration gap is stated as-is. Resolution and compute differ, so this is not an equal-budget comparison or a number to combine with text indices/Pi pass rates.
The same verified video-to-video score, with Custom prompts in our recorded rows. The BoN×8 Cosmos rows carry (Astra) — the board notes GPT-6-Astra prompt upsampling; the unenhanced single-run rows score lower (Super 50.80 / Nano 43.00), and that configuration gap is stated as-is. Resolution and compute differ, so this is not an equal-budget comparison or a number to combine with text indices/Pi pass rates.
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).
NVIDIA-reported benchmarks and boundary statements
The 120B-total, 12B-active multimodal model launches with public weights, training materials and hosted routes; model cards specify precision and hardware.
The 74B-total, 9B-active local-agent model launches. 24GB names a specific quantization recipe; long-context GGUF requires more memory. The October 15 source still described E4B as a later plan; its October 23 release is listed separately.
A specialist continues endoscopic video from two-arm actions; distinguish two-/four-step weights and the GPU runtime. Clinical validity and a matching score in the existing physics board remain unestablished.
The research and card date GA to October 23: Gemma-4-E4B post-training, 8B total/4B effective, with BF16 and Q4_K_M GGUF artifacts. 4GB describes a target VRAM envelope, not total system memory.
A speech-transcription SALM appears on HF trending: 40 languages, recordings to one hour, 34B total / about 4.3B active including the audio encoder, BF16/FP8/NVFP4, servable on one H100, OpenMDW-1.1 license. Card read October 29; the card does not state a weight-release date, so this is our check date.
The official HF blog announces Cosmos3-Nano-Transfer-Auto/Sim-Auto/Sim-Transfer/Sim-Bimanual: 16B trainable parameters, OpenMDW1.1, for 11-view + LiDAR driving simulation and bimanual robot actions to video; repos created August, weights updated Oct 30, blog Nov 5. Physics board gains Cosmos rows, Super (Astra) 65.47 second; AA/Pi do not apply.