NVIDIA Nemotron 3.5 Super VL Is a Builder Just Like You

NVIDIA released Nemotron 3.5 Super VL, an open multimodal model that keeps the 120B total / 12B active parameter profile of Nemotron 3 Super while adding image and video input via a RADIO vision encoder, a vision projector, and a dedicated temporal embedder for video. The model offers a 1M-token context window with up to 131,072 output tokens, toggleable reasoning with reasoning_effort settings of off, medium, or on and a hard reasoning_budget cap, tool calling, and JSON-schema structured output, and Kilo by Anaconda tested it in early access across Kilo Desktop, Kilo CLI, VS Code, and JetBrains.

Nemotron has become a family affair at Kilo by Anaconda. Nemotron 3 Super launched as a 120B, A12B model and went live in Kilo on day one. Nemotron 3 Ultra https://blog.kilo.ai/p/nvidia-nemotron-3-ultra followed as NVIDIA’s flagship, and in August we went hands-on with Nemotron 3.5 Lightning https://blog.kilo.ai/p/nvidia-nemotron-3-5-lightning , the fastest Nemotron yet. Along the way, Nemotron 3 Super became a daily driver for many Kilo Coders. Now NVIDIA is back with what feels like a big leap forward for builders: NVIDIA Nemotron 3.5 Super VL https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/ . We’ve been testing it in early access, and one thing stood out right away. This is a model built for hybrid work : sessions that move between code, data, screenshots, and video, and between Kilo Desktop, the CLI, and the IDE, without losing the thread. What’s New in Nemotron 3.5 Super VL Nemotron 3.5 Super VL keeps the 120B total / 12B active parameter profile that made its predecessor so efficient, but it adds something big: it can see. Under the hood, Nemotron 3.5 Super VL https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning/ pairs a Nemotron-H hybrid language decoder with a RADIO vision encoder, a vision projector, and a dedicated temporal embedder for video, while keeping Multi-Token Prediction weights. That builds on the architecture NVIDIA introduced with Nemotron 3 Super https://blogs.nvidia.com/blog/nemotron-3-super-agentic-ai/ , where Mamba layers handle memory and compute efficiency, transformer layers drive reasoning, and only 12B of 120B parameters are active at inference. Here’s what that means for builders: - A 1M-token context window with up to 131,072 output tokens, enough to hold an entire codebase, a long agentic session, and the docs that go with it. - Image and video input. Pass screenshots, UI mockups, charts, diagrams, or MP4 clips directly into the conversation. - Thinking on demand. Toggle reasoning on or off, set reasoning effort to off, medium, or on, or cap thinking with a hard reasoning budget so you control the speed/cost tradeoff task by task. - Tool calling with auto, required, and named function selection, the backbone of reliable agentic work. - Structured output via JSON schema, so agents return data your pipelines can actually parse. That combination is rare in an open model. You get frontier-style controls over reasoning, native multimodal input, and the efficiency of a 12B-active MoE. Built for Hybrid Work Most real projects don’t stay in one place. And as your real org grows, so do the virtual agent setups to support it. You scaffold in the terminal, iterate in VS Code or JetBrains, explore a dataset in a notebook, then check a screenshot of a broken layout. Every handoff is a chance for context to get lost. That’s where Nemotron 3.5 Super VL has been especially strong for us, because Kilo was built for exactly this kind of work. Kilo Desktop. Kilo Desktop https://blog.kilo.ai/p/desktop puts an agent beside your code and your notebooks. From one conversation, the agent can write and edit code, create Python environments, explore data in notebooks, and review git changes. With vision in the loop, Nemotron 3.5 Super VL can read the chart it just plotted, spot what looks wrong, and fix the cell. Desktop also manages conda environments in-app, drawing from Anaconda’s catalog of more than 19,000 vetted packages. Kilo CLI. In the terminal, Nemotron 3.5 Super VL’s efficiency and tool calling make it a dependable workhorse for scaffolding, refactors, and git-heavy tasks. Dial reasoning effort down for fast execution, or up when the task calls for planning. VS Code and JetBrains. In the IDE, paste a screenshot of a design or a bug and let the agent work from what it sees, not just what you can describe. The best part is that these aren’t separate worlds. Kilo Desktop starts a local copy of the Kilo CLI and drives it through chat, so sessions use the same runtime that powers the CLI and the VS Code extension. Nemotron 3.5 Super VL behaves the same way wherever you pick up the work. Where Nemotron 3.5 Super VL Shines In early testing, we’ve seen Nemotron 3.5 Super VL excel at: - Screenshot-to-code fixes. Reading UI screenshots and making targeted front-end changes. - Data work in notebooks. Interpreting plots and tables, then iterating on the analysis. - Long-context agentic sessions. Holding large repos and long tool-call chains in memory without drifting. - Structured, schema-bound tasks. Returning clean JSON for downstream automation. Nemotron 3.5 Super VL scored 53% in KiloBench , similar to GLM 5.2 also 53% in KiloBench . GLM 5.2 has been one of our most popular models ever, a daily driver across coding tasks and agentic work, so Nemotron 3.5 Super VL’s similarly high score is very promising for builders looking into open models that can be run anywhere and support a wide range of coding and data work. The Right Nemotron for the Right Task With Nemotron 3.5 Super VL, the Nemotron lineup in Kilo covers the full range: - Nemotron 3 Ultra for deep reasoning and orchestration. - Nemotron 3.5 Super VL for efficient, multimodal, hybrid work, and a strong daily driver. - Nemotron 3.5 Lightning as a fast executor for bounded sub-tasks in Agent Manager https://kilo.ai/docs/automate/agent-manager . That’s exactly the point we’ve been making about model freedom https://blog.kilo.ai/p/new-models-from-openai-anthropic-spacexai . The useful question is no longer “which is the best model?” but “which model is best for this task?” Pair Nemotron 3 Ultra as an orchestrator with Nemotron 3.5 Super VL and Nemotron 3.5 Lightning as executors, or let auto models https://kilo.ai/docs/code-with-ai/agents/auto-model handle the routing. NVIDIA Nemotron models have been popular in our auto routing because they offer an ideal balance of cost and performance. Open Models Keep Raising the Bar NVIDIA is treating open models as a core strategy, not a side project. In March, Wired reported https://www.wired.com/story/nvidia-investing-26-billion-open-source-models/ that NVIDIA expected to spend about $26 billion over five years on open-weight models, and the Nemotron Coalition brought labs including Mistral, Perplexity, LangChain, and Thinking Machines together around open frontier models. We’re seeing that commitment in action with a number of new initiatives, the Hugging Face acquisition https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/ , and releases like Nemotron 3.5 Super VL. Each release in the Nemotron line has pushed the open frontier further. Nemotron 3.5 Super VL adds vision and video to that efficiency story, which makes it an even stronger fit for teams that want capable models they can inspect, customize, and deploy on their own terms. That matters for us. Kilo is open source and model-agnostic, and as part of Kilo by Anaconda https://www.anaconda.com/products/kilo-by-anaconda , we’re bringing that same model freedom to enterprise teams, with governance over which models, providers, and agent permissions people can use. Try Nemotron 3.5 Super VL in Kilo Today NVIDIA Nemotron 3.5 Super VL is live in Kilo, across Kilo Desktop https://kilo.ai/docs/desktop , the CLI, and the VS Code and JetBrains extensions. You can browse the whole lineup on our NVIDIA models page https://kilo.ai/models/by/nvidia and track how it stacks up on the Kilo Leaderboard https://kilo.ai/leaderboard . Grab a Kilo Pass https://kilo.ai/pricing/kilo-pass , give it a screenshot, a notebook, or your most exciting repo, and let us know where it fits in your workflow