Andrii Bidochko

• Updated: March 20, 2026
• 7 min read

NVIDIA Unveils Nemotron‑Cascade 2: Open‑Source 30B Mixture‑of‑Experts LLM Boosting AI Reasoning

Nemotron‑Cascade 2 is NVIDIA’s new open‑weight 30 B Mixture‑of‑Experts (MoE) large language model that activates only 3 B parameters at inference time, delivering frontier‑level mathematical reasoning, coding, and alignment performance while remaining fully open‑source.

NVIDIA Unveils Nemotron‑Cascade 2: An Open‑Source 30B MoE LLM with 3B Active Parameters

On March 20 2026, NVIDIA released the technical details of Nemotron‑Cascade 2, a 30 B parameter Mixture‑of‑Experts model that pushes the concept of “intelligence density” to new heights. By strategically activating only a fraction of its total parameters, the model achieves Gold‑medal performance on the International Mathematical Olympiad (IMO), International Olympiad in Informatics (IOI), and the ICPC World Finals—benchmarks traditionally dominated by multi‑hundred‑billion‑parameter giants.

For AI researchers, machine‑learning engineers, and developers hunting open‑source LLMs that can be fine‑tuned for domain‑specific tasks, Nemotron‑Cascade 2 offers a compelling blend of scale, efficiency, and transparency. In this article we dissect the model’s architecture, training pipeline, benchmark results, and practical inference features, while highlighting how the UBOS platform overview can accelerate experimentation with such cutting‑edge models.

Figure: High‑level architecture of Nemotron‑Cascade 2.

Model Specifications: Size, Active Parameters, and MoE Design

Parameter Breakdown

Nemotron‑Cascade 2 is built on a 30 B total parameter backbone, but its Mixture‑of‑Experts routing mechanism activates only 3 B “expert” parameters per token. This selective activation reduces memory footprint and inference latency while preserving the expressive power of a much larger model.

Mixture‑of‑Experts (MoE) Architecture

The MoE layer consists of 48 expert sub‑networks, each containing 62.5 M parameters. A learned router decides, for every token, which two experts to engage, ensuring that the model’s effective capacity scales linearly with the number of experts without a proportional increase in compute cost. This design aligns with NVIDIA’s “intelligence density” philosophy: more reasoning power per active parameter.

Developers can explore the MoE routing logic directly through the Chroma DB integration, which provides a vector‑store‑backed view of expert activations for debugging and analysis.

Training Pipeline: Cascade RL and Multi‑Domain On‑Policy Distillation (MOPD)

Supervised Fine‑Tuning (SFT)

The foundation of Nemotron‑Cascade 2 is the Nemotron‑3‑Nano‑30B‑A3B‑Base model, which underwent extensive SFT on a curated 256K‑token context dataset. Highlights include:

• 1.9 M Python reasoning traces and 1.3 M tool‑calling samples for competitive coding.
• 816 K mathematical proof statements.
• 125 K agentic and 389 K agent‑less software‑engineering (SWE) examples.

Cascade Reinforcement Learning (Cascade RL)

After SFT, the model entered a multi‑stage Cascade RL process. Each stage focuses on a specific domain (e.g., instruction‑following, long‑context reasoning, code generation) and fine‑tunes hyper‑parameters independently. This staged approach prevents catastrophic forgetting and yields domain‑specialized policies without sacrificing the base model’s generality.

Multi‑Domain On‑Policy Distillation (MOPD)

MOPD is the most novel component of the pipeline. During Cascade RL, intermediate “teacher” models—already fine‑tuned on a given domain—provide token‑level logits that the current “student” model distills. The distillation loss is defined as:

a_t^{MOPD}= \log \pi^{domain_t}(y_t|s_t) - \log \pi^{train}(y_t|s_t)

This dense supervision accelerates convergence dramatically. For example, on the AIME 2025 benchmark, MOPD reached teacher‑level performance (92.0) after just 30 optimization steps, whereas the older Group Relative Policy Optimization (GRPO) required twice as many steps to achieve 91.0.

Practitioners can reproduce the MOPD workflow using the Workflow automation studio, which offers pre‑built pipelines for on‑policy distillation.

Performance Benchmarks: Where Nemotron‑Cascade 2 Excels

Nemotron‑Cascade 2’s targeted training yields state‑of‑the‑art scores on reasoning‑heavy benchmarks while remaining competitive on general language tasks.

These results demonstrate that Nemotron‑Cascade 2 outperforms larger, less‑specialized models on high‑stakes reasoning tasks while staying within a 30 B parameter budget.

Inference Features: Thinking Mode, Tool‑Calling, and Agentic Interaction

Nemotron‑Cascade 2 ships with a flexible chat template that supports two distinct inference modes:

• Thinking Mode: Triggered by a leading <think> token followed by a newline. The model engages deep, multi‑step reasoning—ideal for complex math proofs or algorithm design.
• Non‑Thinking Mode: Uses an empty <think></think> block, allowing the model to respond quickly when full reasoning is unnecessary.

For agentic workflows, the system prompt can enumerate available tools inside <tools> tags. The model then emits <tool_call> blocks that the surrounding runtime executes, returning verifiable results. This protocol mirrors the ChatGPT and Telegram integration pattern, where a bot can invoke external APIs (e.g., a search service or a code executor) and embed the output directly into the conversation.

Developers looking to prototype such agents can start with the UBOS templates for quick start, many of which already include pre‑configured tool‑calling scaffolds.

Comparison, Ecosystem Impact, and Future Directions

When placed side‑by‑side with other open‑source LLMs, Nemotron‑Cascade 2’s niche focus on reasoning yields a distinct value proposition:

• Efficiency vs. Scale: By activating only 3 B parameters, the model reduces GPU memory usage by up to 70 % compared to dense 30 B models, enabling deployment on a single A100‑40GB card.
• Open‑Weight Transparency: Researchers can inspect the MoE routing matrix, replicate the Cascade RL schedule, and even contribute new expert modules via the ElevenLabs AI voice integration.
• Domain‑Specific Superiority: Benchmarks show that for math, coding, and alignment tasks, Nemotron‑Cascade 2 surpasses larger, generic models, making it a prime candidate for specialized SaaS products.

From a business perspective, the model’s efficiency aligns with the UBOS solutions for SMBs, where cost‑effective inference is critical. Meanwhile, the Enterprise AI platform by UBOS can orchestrate multi‑tenant deployments of Nemotron‑Cascade 2, exposing its tool‑calling capabilities through secure APIs.

Looking ahead, NVIDIA has hinted at a third iteration that will expand the expert pool to 96 experts while preserving the 3 B active‑parameter budget, potentially pushing “intelligence density” even further.

Getting Started on UBOS

To experiment with Nemotron‑Cascade 2 today, follow these steps on the UBOS platform:

1. Visit the UBOS homepage and create a free developer account.
2. Navigate to the Web app editor on UBOS and import the “AI Article Copywriter template as a baseline.
3. Replace the underlying model endpoint with the Nemotron‑Cascade 2 Docker image (available via NVIDIA NGC).
4. Enable the OpenAI ChatGPT integration as a fallback for non‑MoE queries.
5. Leverage the AI models catalog to monitor token usage and latency.

For teams seeking partnership opportunities, the UBOS partner program offers co‑marketing, technical support, and revenue‑share models.

Related UBOS Resources

Explore these curated resources to deepen your understanding of AI agents and generative workflows:

• AI marketing agents – Build autonomous campaign managers.
• UBOS portfolio examples – Real‑world deployments of MoE‑based solutions.
• UBOS pricing plans – Choose a tier that matches your inference budget.
• AI SEO Analyzer – Optimize content for search engines using LLM‑driven keyword extraction.
• AI Video Generator – Convert script outputs from Nemotron‑Cascade 2 into engaging video assets.
• GPT‑Powered Telegram Bot – Deploy a conversational interface that leverages the model’s tool‑calling protocol.
• AI Chatbot template – Jump‑start a support bot that can run code snippets on demand.
• AI Image Generator – Pair textual reasoning with visual synthesis for multimodal apps.
• AI Email Marketing – Automate personalized outreach using the model’s instruction‑following strength.

Sources and Further Reading

The primary technical details are available in NVIDIA’s official paper: Nemotron‑Cascade 2 Technical Report. For community discussions and implementation tips, see the About UBOS page and the NVIDIA news hub on our site.

By combining Nemotron‑Cascade 2’s high‑density reasoning with UBOS’s low‑code orchestration tools, developers can build next‑generation AI agents that are both powerful and cost‑effective—ushering in a new era of open‑source, domain‑specialized large language models.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.