# NVIDIA presents Nemotron 3.5 models for agentic workloads [Canonical reading](https://escapemeridian.com/readings/nvidia-presents-nemotron-3-5-models-for-agentic-workloads-9023d85c) NVIDIA’s authors introduce Nemotron 3.5 Super and Lightning, open models built for reasoning, long-context understanding, and agentic workloads—tasks where models use tools or take actions. Super is a multimodal model, while Lightning is a smaller text model; both use a Mixture-of-Experts Hybrid Mamba-Transformer design and are intended for efficient deployment. The authors describe training and deployment methods aimed at capability and efficiency. They report that Super reaches 51.78% on SWE-bench Pro, a software-engineering benchmark, and say it surpasses Nemotron 3 Ultra while using approximately 4.6x fewer active parameters. They also report improvements on Indian-language benchmarks, speculative-decoding acceptance length, and inference speed, alongside results for quantization. These are reported evaluations and performance figures, not evidence here of reliability in everyday use. The work matters as an attempt to combine agent-oriented capabilities with lower active parameter counts and faster inference. The abstract does not establish how the models perform across independent evaluations, real-world deployments, or different user needs. How well do the reported benchmark and efficiency results carry over to independent tests and practical agentic workloads? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 28, 2026 (2026-10-28). Date basis: Date basis not provided. Added to Escape Meridian: 2026-10-30T14:18:24.350Z. First observed: 2026-10-30T06:45:32.424Z. Retrieved: 2026-10-30T14:17:36.344Z. - [Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI](https://arxiv.org/abs/2610.38095) ## Limits Evidence is the arXiv abstract, not the full paper; it does not provide evaluation protocols, detailed comparisons, or independent replication. The abstract reports benchmark and deployment figures but does not establish reliability or performance in everyday use or across practical agentic workloads. ## Open question How well do the reported benchmark and efficiency results carry over to independent tests and practical agentic workloads? ## Source-stated The authors introduce two open models designed for reasoning, long-context understanding, and agentic workloads. > We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated Super is a 120B-parameter multimodal model with 12B active parameters; Lightning is a 30B-parameter text model with 3B active parameters. > Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated The authors report Super achieves 51.78% on SWE-bench Pro and surpasses Nemotron 3 Ultra using approximately 4.6x fewer active parameters. > The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated The authors report that Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points. > Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated The abstract reports inference speeds up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user. > Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user. - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source and contributor context NVIDIA, :, Jhunjhunwala, Aastha, Srivastava, Aayushi Ashokkumar, Gupta, Abhibha, Garg, Abhinav, Khattar, Abhinav, Wildish, Achsah, Rajfer, Adam, Renduchintala, Adi, Asif, Adil, Miranda, Adriana Flores, Zaman, Aileen, Chhabra, Aishwarya, Shukla, Akanksha, Bercovich, Akhiad, Rico, Alejandra, Gordic, Aleksa, Ficek, Aleksander, Gronskiy, Alex, Steiner, Alex, Bukharin, Alexander, Milesi, Alexandre, Ghias, Ali Roshan, Shehper, Ali, Taghibakhshi, Ali, Gatti, Alice, Liu, Alisa, Kumar, Alok, Tripathy, Alok, Deshmukh, Amala Sanjay, Miryala, Aman, Shanbhag, Aman, Mahabaleshwarkar, Ameya Sunil, Velingker, Ameya, Klein, Amir, Samani, Amir, Zuker, Amit, Geifman, Amnon, Bhiwandiwalla, Anahita, Subramaniam, Ananth, Manoel, Andre, Santilli, Andrea, Tao, Andrew, Skliar, Andrii, Thomas, Anil, Mahishi, Anish, Shors, Anna, Warno, Anna, Doshi, Anuj, Kamath, Anusha, Kiran, Anwith, Tang, Ao, Mehta, Arham, Dhara, Arnab, Komaragiri, Arnav, Kundu, Arnav, Jain, Arti, Rao, Arun, Venkatesan, Arun, Goel, Arushi, Anoosheh, Asha, Vaswani, Ashish, Aithal, Ashwath, Ahamed, Asif, Thekkumpate, Asma Kuriparambil, Visva, Aswin, Sohrabizadeh, Atefeh, Vem, Avinash, Dattagupta, Ayush, Schifferer, Benedikt, Chislett, Benjamin, Nguyen, Bernard, Nushi, Besmira, Kartal, Bilal, Rouhani, Bita Darvish, Chen, Bobby, Ginsburg, Boris, Wang, Boxin, Norick, Brandon, Kisacanin, Branislav, Johnson, Brendan, Dang, Brian, Yu, Brian, Catanzaro, Bryan, Lee, Byung-Kwan, del Mundo, Carlo, Lee, Chankyu, Hwang, Chantal, Wang, Charles, Truong, Charlie, Cui, Chen, Wang, Cheng, Hsieh, Cheng-Ping, Yu, Chenhan, Luo, Chenjie, Patel, Chintan, Alexiuk, Chris, Johnson, Chris, Wing, Chris · Article authors (partial credit list) These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Agentic AI AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#agentic) ## Multimodal model A model that works with more than one kind of input or output, such as text, images and audio. - [Definition reference](https://developers.google.com/machine-learning/glossary#multimodal-model) ## Mixture of experts A model design that sends each input to a few specialized sub-networks instead of the whole model, which saves computation. - [Definition reference](https://developers.google.com/machine-learning/glossary#mixture-of-experts) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Quantization Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy. - [Definition reference](https://developers.google.com/machine-learning/glossary#quantization) ## Model parameters The internal numbers a model adjusts during training; their count is a rough measure of a model's size. - [Definition reference](https://developers.google.com/machine-learning/glossary#parameter) ## NVIDIA NVIDIA develops processors, networking and software used for AI, graphics and scientific computing. - [Explore this context](/entities/nvidia) - [Reference](https://blogs.nvidia.com/) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 210ed1d2-369f-4613-871e-5585f4f35d9d. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-10-30T14:18:24.350Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: c53949a5a18cc1554ea2a08b215a502c2e84d1b9d970cb422b0fe54a5986946c. Source text hash: ba849d11c8da60be17f24560a485c35f417f762b1053826fb466008c64925881. ## Revision and source-check history ### Version 1 — 2026-10-30T14:18:24.350Z NVIDIA’s authors introduce Nemotron 3.5 Super and Lightning, open models built for reasoning, long-context understanding, and agentic workloads—tasks where models use tools or take actions. Super is a multimodal model, while Lightning is a smaller text model; both use a Mixture-of-Experts Hybrid Mamba-Transformer design and are intended for efficient deployment. The authors describe training and deployment methods aimed at capability and efficiency. They report that Super reaches 51.78% on SWE-bench Pro, a software-engineering benchmark, and say it surpasses Nemotron 3 Ultra while using approximately 4.6x fewer active parameters. They also report improvements on Indian-language benchmarks, speculative-decoding acceptance length, and inference speed, alongside results for quantization. These are reported evaluations and performance figures, not evidence here of reliability in everyday use. The work matters as an attempt to combine agent-oriented capabilities with lower active parameter counts and faster inference. The abstract does not establish how the models perform across independent evaluations, real-world deployments, or different user needs. How well do the reported benchmark and efficiency results carry over to independent tests and practical agentic workloads? ## Original source arXiv cs.AI · Article Original source date: Oct 28, 2026 (2026-10-28). Date basis: Date basis not provided. Added to Escape Meridian: 2026-10-30T14:18:24.350Z. First observed: 2026-10-30T06:45:32.424Z. Retrieved: 2026-10-30T14:17:36.344Z. - [Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI](https://arxiv.org/abs/2610.38095) ## Limits Evidence is the arXiv abstract, not the full paper; it does not provide evaluation protocols, detailed comparisons, or independent replication. The abstract reports benchmark and deployment figures but does not establish reliability or performance in everyday use or across practical agentic workloads. ## Open question How well do the reported benchmark and efficiency results carry over to independent tests and practical agentic workloads? ## Source-stated The authors introduce two open models designed for reasoning, long-context understanding, and agentic workloads. > We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated Super is a 120B-parameter multimodal model with 12B active parameters; Lightning is a 30B-parameter text model with 3B active parameters. > Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated The authors report Super achieves 51.78% on SWE-bench Pro and surpasses Nemotron 3 Ultra using approximately 4.6x fewer active parameters. > The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated The authors report that Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points. > Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source-stated The abstract reports inference speeds up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user. > Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user. - [Supporting source](https://arxiv.org/abs/2610.38095) ## Source and contributor context NVIDIA, :, Jhunjhunwala, Aastha, Srivastava, Aayushi Ashokkumar, Gupta, Abhibha, Garg, Abhinav, Khattar, Abhinav, Wildish, Achsah, Rajfer, Adam, Renduchintala, Adi, Asif, Adil, Miranda, Adriana Flores, Zaman, Aileen, Chhabra, Aishwarya, Shukla, Akanksha, Bercovich, Akhiad, Rico, Alejandra, Gordic, Aleksa, Ficek, Aleksander, Gronskiy, Alex, Steiner, Alex, Bukharin, Alexander, Milesi, Alexandre, Ghias, Ali Roshan, Shehper, Ali, Taghibakhshi, Ali, Gatti, Alice, Liu, Alisa, Kumar, Alok, Tripathy, Alok, Deshmukh, Amala Sanjay, Miryala, Aman, Shanbhag, Aman, Mahabaleshwarkar, Ameya Sunil, Velingker, Ameya, Klein, Amir, Samani, Amir, Zuker, Amit, Geifman, Amnon, Bhiwandiwalla, Anahita, Subramaniam, Ananth, Manoel, Andre, Santilli, Andrea, Tao, Andrew, Skliar, Andrii, Thomas, Anil, Mahishi, Anish, Shors, Anna, Warno, Anna, Doshi, Anuj, Kamath, Anusha, Kiran, Anwith, Tang, Ao, Mehta, Arham, Dhara, Arnab, Komaragiri, Arnav, Kundu, Arnav, Jain, Arti, Rao, Arun, Venkatesan, Arun, Goel, Arushi, Anoosheh, Asha, Vaswani, Ashish, Aithal, Ashwath, Ahamed, Asif, Thekkumpate, Asma Kuriparambil, Visva, Aswin, Sohrabizadeh, Atefeh, Vem, Avinash, Dattagupta, Ayush, Schifferer, Benedikt, Chislett, Benjamin, Nguyen, Bernard, Nushi, Besmira, Kartal, Bilal, Rouhani, Bita Darvish, Chen, Bobby, Ginsburg, Boris, Wang, Boxin, Norick, Brandon, Kisacanin, Branislav, Johnson, Brendan, Dang, Brian, Yu, Brian, Catanzaro, Bryan, Lee, Byung-Kwan, del Mundo, Carlo, Lee, Chankyu, Hwang, Chantal, Wang, Charles, Truong, Charlie, Cui, Chen, Wang, Cheng, Hsieh, Cheng-Ping, Yu, Chenhan, Luo, Chenjie, Patel, Chintan, Alexiuk, Chris, Johnson, Chris, Wing, Chris · Article authors (partial credit list) These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Agentic AI AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#agentic) ## Multimodal model A model that works with more than one kind of input or output, such as text, images and audio. - [Definition reference](https://developers.google.com/machine-learning/glossary#multimodal-model) ## Mixture of experts A model design that sends each input to a few specialized sub-networks instead of the whole model, which saves computation. - [Definition reference](https://developers.google.com/machine-learning/glossary#mixture-of-experts) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Quantization Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy. - [Definition reference](https://developers.google.com/machine-learning/glossary#quantization) ## Model parameters The internal numbers a model adjusts during training; their count is a rough measure of a model's size. - [Definition reference](https://developers.google.com/machine-learning/glossary#parameter) ## NVIDIA NVIDIA develops processors, networking and software used for AI, graphics and scientific computing. - [Explore this context](/entities/nvidia) - [Reference](https://blogs.nvidia.com/) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 210ed1d2-369f-4613-871e-5585f4f35d9d. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-10-30T14:18:24.350Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: c53949a5a18cc1554ea2a08b215a502c2e84d1b9d970cb422b0fe54a5986946c. Source text hash: ba849d11c8da60be17f24560a485c35f417f762b1053826fb466008c64925881. ### passed — 2026-10-30T14:18:24.350Z — anthropic/claude-opus-5.5 Publication: 210ed1d2-369f-4613-871e-5585f4f35d9d attribution — pass: The draft attributes the work and its reported results to NVIDIA's authors. Benchmark and efficiency figures are framed as reported by the authors, not as established fact. > Authors:NVIDIA: Aastha Jhunjhunwala summary — pass: The title, summary and question match the abstract. The draft names the actor and the two models. It explains agentic workloads and SWE-bench Pro in plain language and states the efficiency aim. Technical terms such as Mixture-of-Experts Hybrid Mamba-Transformer, active parameters and speculative decoding are kept exactly, suited to tooltips. The question is a genuine open issue that follows from the abstract-only evidence. > Across training and deployment, we jointly optimize capability and efficiency. date — pass: The originalPublished date of 2026-10-28 matches the arXiv v1 submission date. It is not the later feed date of 2026-10-30. > \[Submitted on 28 Oct 2026\] support — pass: Each claim is directly supported by an exact abstract quote: the parameter counts, the 51.78% SWE-bench Pro score, the 4.6x comparison, the up to 9.82-point Indic gain and the throughput figures. The significance sentence about combining capability with efficiency reflects the authors' stated aim. The caveats correctly avoid turning benchmark scores into claims of real-world reliability. > The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. media — pass: The selected media is the verified direct link to the original arXiv abstract page. It is the most direct way to access the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/9023d85cd47bbed3cd6d6804d40257cc659284de7d57c0284c89b2a64bcf1e64/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/nvidia-presents-nemotron-3-5-models-for-agentic-workloads-9023d85c.json)