Written by Jakub Rusinowski · Last updated July 21, 2026
Model library → Cosmos 3 → Cosmos 3 Edge
The on-device tier — a 4B-parameter open world model on a dense 2B backbone that runs vision reasoning and robot-action generation directly on the robot, with no cloud round-trip. Released mid-July 2026, Edge targets NVIDIA Jetson (including the new Jetson T2000/T3000 modules) and RTX GPUs; NVIDIA cites a GeForce RTX 3070 or better as a local prototyping on-ramp. On Jetson Thor it generates 32 actions per inference at 15 Hz real-time control. Weights are ~8 GB at BF16, so it fits a single consumer or embedded GPU. This is the practical tier for hobbyist and edge robotics. Omnimodal reasoning + action. OpenMDW-1.1. Specs from launch coverage and the Hugging Face model card — verify before relying on them.
Cosmos 3 Edge needs about 3 GB of VRAM at BF16 — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
| Parameters | 4B total (2B dense transformer backbone) |
| Context window | Omnimodal I/O (video/action) |
| Architecture | Two-tower Mixture-of-Transformers — autoregressive reasoner (VLM) + diffusion generator, shared multimodal attention |
| Provider | NVIDIA |
| Licence | OpenMDW-1.1 |
| Specified at | BF16 |
| System RAM | 16 GB |
| Record updated | 2026-07-21 |
OpenMDW-1.1 — commercial use permitted. Commercial use permitted. No usage restrictions beyond attribution.
Modelled on a reference NVIDIA RTX 4090 (24 GB), with no KV cache (this record has no published architecture). Speed figures are ESTIMATES from the memory-bandwidth roofline described on the methodology page, not benchmarks we ran — rows marked measured come from published or reader-submitted runs. VRAM here includes the KV cache, so it reads higher than the headline figure above, which does not.
| Quant | Weights | VRAM needed | Est. speed | Fit on 24 GB |
|---|---|---|---|---|
| Q2_K | 1.3 GB | 2.1 GB | ~269 tok/s (est.) | Fits comfortably |
| Q3_K_M | 1.7 GB | 2.5 GB | ~252 tok/s (est.) | Fits comfortably |
| Q4_K_M | 2.4 GB | 3.2 GB | ~225 tok/s (est.) | Fits comfortably |
| Q5_K_M | 2.8 GB | 3.6 GB | ~212 tok/s (est.) | Fits comfortably |
| Q6_K | 3.3 GB | 4.1 GB | ~199 tok/s (est.) | Fits comfortably |
| Q8_0 | 4.3 GB | 5.0 GB | ~177 tok/s (est.) | Fits comfortably |
| F16 | 8.0 GB | 8.8 GB | ~123 tok/s (est.) | Fits comfortably |
Want the memory numbers alone, at every quantization level and your own context length? Use the Cosmos 3 Edge VRAM calculator.
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
The cheapest catalogued GPU that runs Cosmos 3 Edge is the Intel Arc B570 (10 GB).
Install Ollama, then run:
ollama run cosmos-3Weights on Hugging Face: nvidia/Cosmos3-Edge.
Best for: edge robotics, on device, real time control, jetson, prototyping.
← All Cosmos 3 models | VRAM calculator | Check your own hardware