NVIDIA's open-weight Nemotron 3.5 Super VL vision-language model tested on bug fixing, translation, circuit diagrams, and creative coding.
What is Nemotron 3.5 Super VL? #
Nemotron 3.5 Super VL is an open-weight vision-language model from NVIDIA, released as a 120 billion parameter model that activates only 12 billion parameters per forward pass. It takes text, images, and video as input and produces text output, with a context window that stretches up to a million tokens. NVIDIA built it for agentic workloads: tasks that mix reasoning, tool use, and visual understanding in one continuous session rather than a single prompt-and-response exchange. The model is available on Hugging Face in formats including NVFP4 and BF16, and ships with its own reasoning parser and chat template for integration into agent frameworks.
TL;DR #
-
Mixture-of-experts architecture keeps the model fast despite its size: 120 billion total parameters but only 12 billion active at inference time, which NVIDIA pairs with Mamba 2 and attention layers rather than a pure transformer stack.
-
Agentic debugging worked in a real test : paired with the Hermes agent framework, the model found and fixed a genuine pricing bug in a prescription lens ordering app without being told what was wrong, though it also made three unnecessary changes to files that weren’t broken.
-
Translation is strong in major languages, weak in low-resource ones : output across English, Spanish, French, and similar languages followed formatting instructions cleanly, but results degraded badly for Burmese, Pashto, Sindhi, and several African languages.
-
Scientific reasoning held up on a hand-drawn circuit : the model read handwritten component values on a Sallen-Key filter, calculated the correct cutoff frequency, and caught a real wiring error in the op-amp inputs, though its written explanation of the error was slightly imprecise.
-
Creative, code-only image generation was the weakest result : asked to build a dragon tattoo design in pure HTML and CSS, the model nailed thematic elements like a folded flag and dog tags but failed to render anything resembling a dragon.
-
Benchmark standing is mixed : NVIDIA’s own comparisons show it trailing models like Qwen3.8 Flash, GLM 5.2 Flash, and DeepSeek on agentic and coding tests, while leading on long-context tasks and running roughly twice as fast as rivals on NVIDIA GPUs.
-
✕a coding agent
-
✕no-code
-
✕vibe coding
-
✕a faster Cursor
The one that tells the coding agents what to build.
How does Nemotron 3.5 Super VL’s architecture work? #
The model uses a mixture-of-experts design, which means not every parameter fires on every token. Out of 120 billion total parameters, only about 12 billion are active per inference step, which is what keeps response times fast relative to the model’s stated size. NVIDIA combines this with Mamba 2 state-space layers and traditional attention layers in the same architecture, a hybrid approach aimed at balancing long-context efficiency with the pattern-matching strength attention layers provide.
On the input side, the model handles text, images, and video, and the vision tower is built around a component called RADIO for image processing. The context window goes up to a million tokens, which in practice means it can hold a large code base, a long document, or an extended multi-turn agent conversation without needing to truncate history. The model card lists support for English, Spanish, French, German, Italian, and Japanese as primary languages, though the hands-on test pushed translation into dozens of additional languages beyond that core set.
How did it perform on agentic bug-fixing? #
In a hands-on test, the model was connected to the Hermes agent framework and pointed at a locally running web application for ordering prescription sunglasses, built with a typical stack of containers for the frontend, backend, and a Redis cache. The app had a real, subtle bug: two patients with identical prescriptions, one written in plus-cylinder notation and one in minus-cylinder notation, were ending up with different lens thickness, different frame compatibility, and different prices. Nothing crashed. The bug was a quiet logic error, the kind that’s easy to miss in code review but costly in production since customers would be overcharged or given the wrong prescription strength entirely.
Given no hints about what was wrong, the model used tool access to inspect the code, found the actual defect (a missing sphere value update in the plus-to-minus cylinder conversion logic), patched it, rebuilt the container, and verified the fix worked by checking that both patients now produced matching, correct lab sheets. That’s a nontrivial agentic loop: read code, reason about domain logic it wasn’t told about, write a fix, rebuild, and confirm. The one flaw was overcorrection. The model also reported fixing three other issues in unrelated files that weren’t actually broken, a sign it may be prone to touching more code than necessary once it’s in an editing mindset.
Is it good at translation and multilingual tasks? #
Given an image of Portuguese text, the model was asked to read it, extract the core meaning, and generate a three-word title in 82 different languages, including several low-resource ones. It correctly identified the central idea of the passage (a theme of contentment over happiness) and produced clean, properly formatted titles in English, Spanish, French, and other widely spoken languages, sticking to the one-language-per-line format requested without extra commentary.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
The results fell apart for less common languages. Burmese output was described as unreadable garbage, Malayalam output had an English word mixed in where it shouldn’t have been, and Pashto and Sindhi outputs were unusable. Several African language outputs were also off. This pattern, strong performance on major world languages and weak performance on low-resource ones, is common across most current large language models, so Nemotron 3.5 Super VL isn’t an outlier here, just not a breakthrough either.
Can it handle scientific and visual reasoning? #
A hand-drawn Sallen-Key filter circuit diagram, the kind of thing an engineering student might sketch by hand, tested three separate skills at once: reading handwritten component values, performing the actual filter math, and catching errors in the circuit’s wiring. The model read every labeled value correctly, correctly identified the circuit topology, and calculated a cutoff frequency around 750 Hz that matched independent verification. More impressively, it went beyond just computing numbers and checked whether the drawn circuit would actually function as intended, catching that the op-amp’s input pins were wired backward, which would prevent the filter from working as drawn.
The explanation of exactly why the wiring was wrong was slightly imprecise (it described a missing feedback connection to the wrong input pin), but the core catch, that the circuit as drawn was non-functional, was correct and delivered quickly without extended reasoning time. For a task that blends visual reading, math, and conceptual engineering judgment, that’s a solid result.
How does it compare to other open models? #
NVIDIA’s own benchmark comparisons place Nemotron 3.5 Super VL below several current rivals on agentic and coding tests, including Qwen3.8 Flash, GLM 5.2 Flash, and DeepSeek, while it does outperform the older Qwen3.5. Its clearer advantage shows up in long-context tasks, where it leads the comparison group, and in raw inference speed on NVIDIA hardware, where it runs at roughly double the token throughput of comparable models. In agent-tool benchmarks, the Hermes framework paired with this model landed second behind Codex and ahead of Claude Code and SWE-bench Pro, in a field where most competitors cluster tightly together.
The creative coding test exposed a clear limitation. Asked to build a dragon tattoo design using nothing but HTML and CSS (no image generation, just code-drawn shapes), the model nailed the surrounding symbolism: a folded flag with correct colors, dog tags, a glowing sunrise, and a memorial banner. But the dragon itself, the actual subject of the piece, didn’t read as a dragon at all: no visible claws, no coiling body, nothing distinctly reptilian. For a 120 billion parameter model, that’s a meaningful gap between understanding a creative brief and executing the hardest visual element of it in pure code.
Frequently Asked Questions #
What is Nemotron 3.5 Super VL used for?
It’s designed for agentic tasks that combine visual understanding, reasoning, and tool use, such as debugging applications, reading diagrams or handwritten notes, and handling long documents or codebases within its million-token context window.
How many parameters does Nemotron 3.5 Super VL have?
It has 120 billion total parameters but uses a mixture-of-experts design that activates only about 12 billion parameters per inference step, which keeps it faster than a dense model of equivalent size.
Is Nemotron 3.5 Super VL open source?
Yes, NVIDIA released it with open weights on Hugging Face, available in formats including NVFP4 and BF16, under NVIDIA’s own license terms.
How does it compare to models like Qwen and DeepSeek?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
On NVIDIA’s own agentic and coding benchmarks, it trails models like Qwen3.8 Flash, GLM 5.2 Flash, and DeepSeek, but it leads on long-context tasks and runs roughly twice as fast on NVIDIA GPUs.
What are its weaknesses?
Testing showed it struggles with low-resource languages in translation tasks, tends to over-edit code when fixing bugs (changing things that weren’t broken), and has difficulty rendering specific visual subjects accurately when generating code-based graphics.