Model Overview & Capabilities
Nemotron 3.5 Super VL is NVIDIA’s open-weights 120B Mixture-of-Experts architecture (activating 12 billion parameters per token) officially released on October 14, 2026, holding Rank #24 on the Vibe Coder Journal Leaderboard with a composite index score of 34.2. Superseding the earlier Nemotron 3 Ultra architecture, it scores 34.5% on SWE-bench Verified, 34.0% on Terminal-Bench 2.0, 35.0% on Aider Polyglot, and 72.0% on GPQA Diamond scientific evaluations.
Available under the permissive NVIDIA Open Model License with hosted API rates at $0.50 input and $1.50 output per million tokens, Nemotron 3.5 delivers a blended price of $0.75 per million tokens under 3:1 weighting, providing a solid cost efficiency of $0.0219 per index point while generating outputs at 137 tokens per second.
Optimized natively for NVIDIA Blackwell and Hopper tensor cores with an extensive 1,000,000-token context window, Nemotron 3.5 Super VL is engineered for GPU-accelerated agent pipelines, automated hardware driver debugging, visual software inspection, CUDA kernel optimization, and high-performance local AI cluster infrastructure. Its specialized vision encoder allows engineers to analyze graphical user interface bug reports, complex system diagrams, and hardware architecture schematics seamlessly alongside source code files, making it a versatile tool for full-stack robotics and systems software engineering.
