Nemotron 3.5 Super
A 120B-parameter multimodal large model from NVIDIA with 12B activated parameters, using a MoE and hybrid Mamba-Transformer architecture, targeting reasoning, long-context understanding, and agentic workloads.
Timeline (last 90 days)
Both models use a mixture-of-experts (MoE) and hybrid Mamba-Transformer architecture, targeting reasoning, long-context understanding and agentic workloads.
Super is a 120B-parameter, 12B-activated-parameter multimodal model; Lightning is a 30B-parameter, 3B-activated-parameter text-only model.
The two models share the same training philosophy: large-scale distillation, long-context training, and native speculative decoding.
Super additionally layers on Indic continued pretraining and agentic software engineering mid-training.
Super introduces multi-teacher on-policy distillation (MOPD), transferring the base model's capabilities to agentic behavior.
Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points.
After post-training, Super reaches 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra with roughly 1/4.6 the activated parameters of the latter.
Super's multimodal and agentic capabilities depend heavily on its post-training pipeline, and there is currently no public evidence on whether community fine-tuning can reproduce these gains.
All Facts (8)
Both models use a mixture-of-experts (MoE) and hybrid Mamba-Transformer architecture, targeting reasoning, long-context understanding and agentic workloads.
50%UnverifiedSuper is a 120B-parameter, 12B-activated-parameter multimodal model; Lightning is a 30B-parameter, 3B-activated-parameter text-only model.
50%UnverifiedThe two models share the same training philosophy: large-scale distillation, long-context training, and native speculative decoding.
50%UnverifiedSuper additionally layers on Indic continued pretraining and agentic software engineering mid-training.
50%UnverifiedSuper introduces multi-teacher on-policy distillation (MOPD), transferring the base model's capabilities to agentic behavior.
50%UnverifiedIndic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points.
50%UnverifiedAfter post-training, Super reaches 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra with roughly 1/4.6 the activated parameters of the latter.
50%UnverifiedSuper's multimodal and agentic capabilities depend heavily on its post-training pipeline, and there is currently no public evidence on whether community fine-tuning can reproduce these gains.
50%