Nemotron 3 Super, Qwen3.5-122B battle on-prem

This is less a model-quality question than a hardware-fit question. Nemotron is the flashier long-context, speed-first bet, but Qwen still feels like the safer default for most local coding/chat setups.

• –NVIDIA's Nemotron 3 Super is a 120B model with 12B active parameters and up to 1M context, and NVIDIA claims 7.5x throughput versus Qwen3.5-122B. That win is benchmark-specific, though, and the NIM docs peg the model to 8x H100-80GB hardware. [Nemotron page](https://research.nvidia.com/labs/nemotron/Nemotron-3-Super/) [NIM reference](https://docs.api.nvidia.com/nim/reference/nvidia-nemotron-3-super-120b-a12b)
• –Qwen3.5-122B-A10B is also a 122B/10B-active MoE, but it is multimodal and slots into common local stacks like Transformers, vLLM, SGLang, and KTransformers. Its official card shows 262k native context, extension to 1,010,000 tokens, and solid coding scores like 72.0 on SWE-bench Verified and 49.4 on Terminal Bench 2. [Qwen model card](https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8)
• –For coding and chat, that makes Qwen the lower-friction pick if you want one model that can also handle vision later. Nemotron only really wins if you can exploit the long-context and NVIDIA-specific acceleration.
• –The Reddit replies mirror the split: some users are excited about Nemotron's speed/context, while others stay with Qwen because it already works well on mixed local hardware and feels less NVIDIA-specific. [Reddit thread](https://www.reddit.com/r/LocalLLaMA/comments/1s2ounq/nemotron_super_3_vs_qwen35_122b_for_onprem/)