Reka Rho-1 vs Nemotron 3.5 Super VL 120B

At a glance

Reka Rho-1 and Nemotron 3.5 Super VL 120B both process text, images, and video in a single model, but they differ in architecture and maturity. Rho-1 is a compact 19B research preview designed to unify reasoning and robotic control within one token stream. Nemotron 3.5 Super VL is a larger 120B mixture-of-experts model built for enterprise agent workflows and available for immediate deployment and customization.

Factor	Reka Rho-1	Nemotron 3.5 Super VL 120B
Maker	Reka Labs	NVIDIA
Release	October 5, 2026 (research preview)	October 15, 2026
Size	19B parameters	120B parameters (12B active per token)
Architecture	Unified omni-reasoning	Mixture-of-experts
Access	Research preview	Open weights, available now
Pricing	Not published	Not published
Robotics support	Yes (native robotic actions as tokens)	Not published
Best for	Unified multimodal reasoning, video generation, robot control	Enterprise agent workflows, document intelligence, video understanding

Reka Rho-1: strengths and limits

Rho-1’s core architecture collapses separate modality pipelines into one. Rather than routing tasks between specialized text, vision, and robotics models, Rho-1 treats text, images, video, and robotic actions as unified tokens within a single context window. This removes the handoff delays that occur when modular systems pass information between components.

The model demonstrates integrated reasoning across content generation and editing. In Reka’s demonstration, Rho-1 generated an aerial lighthouse image, annotated it with a bounding box, animated the image into video with camera movement, edited the video to add a snowstorm effect, and then answered questions comparing the two videos. This shows the model can reason over generated and edited content without separate pipelines.

Robotic actions are native to the tokenization, allowing the model to reason about and control physical systems without translating between modalities.

The key limitation is maturity and public transparency. Rho-1 is released as a research preview only, not a production system. No independent benchmarks have been published. Pricing is not available. Context window size is not disclosed. Organizations cannot evaluate fit or performance on their own data beyond public demonstrations.

The 19B parameter size is smaller than many current production models, which may constrain reasoning depth on complex tasks, though published benchmarks do not confirm this.

Nemotron 3.5 Super VL 120B: strengths and limits

Nemotron 3.5 Super VL is engineered for production deployment and domain-specific customization. The 120B-parameter mixture-of-experts architecture activates only 12B parameters per token, balancing model capacity with inference efficiency. NVIDIA states the model delivers up to 5x higher throughput for agentic tasks compared to other open models in its class, based on the Artificial Analysis Intelligence Index agentic leaderboard.

The model is built for agent workflows that combine reasoning, tool use, and visual understanding across coding tasks (validation, error analysis), web research (searching, connecting information), document intelligence (extracting from tables and charts), cybersecurity (identifying vulnerabilities), and video understanding (search, content review, incident analysis).

Customization is central to the architecture. Organizations can post-train Nemotron 3.5 Super VL using NVIDIA NeMo with supervised fine-tuning, low-rank adaptation, and reinforcement learning. Organizations including Scale AI, Ramp, AI21, and Mercor have customized the model for domain-specific tasks. Reported improvements from post-training include 40x lower inference cost, 10x lower latency, and 40% higher accuracy on specialized tasks.

The model is open-source with weights published and deployment guidance available via GitHub and NVIDIA NIM.

Limitations include size and infrastructure requirements. At 120B parameters, the model requires more GPU memory than smaller alternatives. While it activates 12B parameters per token, the full model footprint affects infrastructure planning. No published benchmarks isolate performance on robotics tasks, and the focus on enterprise agent workflows means research or consumer applications may find fewer tailored examples.

The throughput advantage is measured against other open models in its size category, so direct performance comparisons to Rho-1 are not available from NVIDIA’s published claims.

Price and access

Reka Rho-1 pricing is not published. The model is available as a research preview with no public information about access mechanisms or timelines.

Nemotron 3.5 Super VL 120B pricing varies by deployment platform. Third-party API listings show input pricing ranging from $0.06 to $0.20 per million tokens and output pricing from $0.35 to $1.00 per million tokens, but these are not official NVIDIA prices. The model is available for self-hosted deployment via open-source weights on Hugging Face, which eliminates per-token costs for organizations running their own infrastructure.

Neither company has published official pricing.

Which one should you choose

Choose Reka Rho-1 if you need a compact model that handles text, image, video generation, and robotic control in a single forward pass without multi-step handoffs. This is best for robotics research, embodied AI prototypes, and video editing workflows where unified reasoning across all modalities is critical. Understand that you would be working with a research preview, so production stability and long-term support are unclear.

Choose Nemotron 3.5 Super VL 120B if you are deploying specialized agent workflows in enterprises, need to customize the model on proprietary data, or require on-premises deployment with full control. This is better for document intelligence, coding assistance, security analysis, and web research where tool use and domain adaptation are essential. If throughput matters and you can run your own infrastructure, the open-source availability and customization tools make this a production-ready choice.

If you need both unified reasoning and enterprise-grade stability, neither is ideal yet. Rho-1 emphasizes unification but lacks production maturity. Nemotron 3.5 Super VL emphasizes agent workflows and customization but does not emphasize unified token sequences the way Rho-1 does.

Consider your latency requirements. Rho-1’s single-model approach theoretically avoids modality handoff delays, but this is not quantified against Nemotron’s throughput. For agents that must complete tasks fast at scale, Nemotron’s 5x throughput advantage on agentic tasks may outweigh Rho-1’s architectural simplicity.

Common questions

What is a mixture-of-experts architecture?

A mixture-of-experts (MoE) model routes different inputs to different specialized subnetworks. Nemotron 3.5 Super VL’s 120B parameters activate only 12B per token, reducing computation while maintaining capacity.

Can I run these models locally?

Nemotron 3.5 Super VL weights are available on Hugging Face for self-hosted deployment. Rho-1 is a research preview with no public download or deployment information.

Which model is faster?

NVIDIA reports Nemotron 3.5 Super VL achieves up to 5x higher throughput for agentic tasks versus other open models in its class. Rho-1’s latency has not been publicly measured.

The conclusion

Reka Rho-1 and Nemotron 3.5 Super VL represent different engineering philosophies. Rho-1 unifies modalities into a single 19B network, eliminating the stack to reduce latency and reasoning fragmentation. Nemotron 3.5 Super VL scales to 120B parameters with efficient mixture-of-experts routing, prioritizing enterprise agent customization and throughput.

For robotics and video generation, Rho-1 is the more direct choice, though it remains unproven outside Reka’s public demonstrations. For enterprise agent deployment, document processing, and specialized workflows, Nemotron 3.5 Super VL is available now with documented performance claims and a track record of successful customization.

Both models are recent October 2026 releases, so neither has extensive real-world deployment history. Evaluate both on your readiness to work with research-preview tools versus production-ready systems, your infrastructure constraints, and whether you need customization on proprietary data. Test both on your specific workflows before committing to either.

Independent human benchmarks comparing these models on robotics, video editing, and agentic tasks do not yet exist in published form. When they become available, revisit this comparison to validate vendor claims against independent evidence.

Related reading

• Synthesia vs Colossyan: AI Training Video
• d3-ultra 27B multimodal decision model scores 87.2
• OpenMobile-2 smartphone agents match proprietary rivals