All models › Compare
DeepSeek-V4-Pro vs NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4)
DeepSeek-V4-Pro and NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) are about even on Arena (1451 vs 1445): in a blind vote each wins about half the time.
DeepSeek-V4-Pro1451Arena chat · #49 of 417
NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4)1445Arena chat · #63 of 417
Arena scores by category
| Category | DeepSeek-V4-Pro | NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) | Ahead |
|---|---|---|---|
| Chat | 1451 | 1445 | about even |
Scores from blind human votes on Arena. A gap under 10 points is too small to call.
Size and where to use them free
| DeepSeek-V4-Pro | NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) | |
|---|---|---|
| Made by | deepseek | nvidia |
| Size | 1598.8 B (75.6 B active) | not public |
| 4-bit download | 970 GB | — |
| Use it free | download and run at home · MIT license | OpenMDW-1.1 license |
Speed at home
| Hardware | DeepSeek-V4-Pro | NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) |
|---|---|---|
| RTX 3090 | does not fit | online only |
| RTX 4090 | does not fit | online only |
| RTX 5090 | does not fit | online only |
| RX 7900 XTX | does not fit | online only |
| Mac M4 Max 128 GB | does not fit | online only |
| Ryzen AI Max+ 395 128 GB | does not fit | online only |
Estimates at 4-bit with an 8K context, 64 GB of system RAM for offload, calibrated on 44 real llama.cpp tests.
Try them yourself
Ask both the same question side by side in Try & compare, or vote blind between free models in Free Arena.
More comparisons
- DeepSeek-V4-Pro vs MiMo-V2.6-Pro
- DeepSeek-V4-Pro vs Kimi K3 (Max)
- DeepSeek-V4-Pro vs GLM-5.3 (Max)
- DeepSeek-V4-Pro vs GLM-5.3-Flash
- DeepSeek-V4-Pro vs GLM-5.2 (Max)
- DeepSeek-V4-Pro vs MiMo-V2.5-Pro
- NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) vs MiMo-V2.6-Pro
- NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) vs Kimi K3 (Max)
- NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) vs GLM-5.3 (Max)
- NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) vs GLM-5.3-Flash
Share this page:XLinkedInRedditWhatsApp
Spotted a wrong number or a broken link? Report a mistake