Unverified50% confidenceFactExact time
Nemotron-3.5-Super-VL comes with a 1.3GB vision tower (C-RADIO) that needs to be streamed layer by layer to the GPU, taking 0.4-1.6 seconds to process each image.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen running Nemotron Super (approximately 120 billion parameters), the unoptimized speed is about 16.6 tokens per second; after enabling MTP, Flash Attention, and KV cache, it improves to 27.2 tokens per second (at a context length of around 17,000), a speedup of approximately 64%.72% similarUnverifiedA single expert of Nemotron-3.5-Super (up 2688×1024 + down 1024×2688) occupies only 3.10MB in NVFP4 format.70% similarUnverifiedThe Nemotron 3.5 Super VL repository is 75.1 GB in size, spread across 88 safetensors shards, and a single consumer-grade card cannot accommodate a 120B model.68% similarUnverifiedNVIDIA claims that the weights of Nemotron 3.5 Bolt 74B-A9B can run in 24GB of memory, and that 24GB refers only to the weights themselves, excluding context, KV cache and runtime overhead.67% similarUnverifiedThe RTX 5090 has only 32 GB of VRAM, which is not enough for Nemotron 3.5 Super VL; even the UD-IQ1_M quantized version requires 46.1 GB.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1053044API
curl https://kongchang.com/api/v1/knowledge/claims/1053044MCP
get_claim(id=1053044)