Nemotron Cascade 2 30B A3B
NVIDIA's open 30B MoE that punches at frontier scale — gold-medal math, coding, and agentic reasoning with only 3B active parameters.
Overview
What is Nemotron Cascade 2 30B A3B
Nemotron Cascade 2 30B A3B is NVIDIA's open-weight Mixture-of-Experts text model released in March 2026. It activates 3B parameters per token from 30B total, delivers gold-medal math and coding reasoning, supports tool use and web search, and runs in both thinking and instruct modes.
Running it privately on Venice
On Venice, Nemotron Cascade 2 runs under a private, zero-retention privacy tier — your prompts are not stored or profiled. You get its reasoning, tool-use, and web-search capabilities with fully open weights, making it a permissionless alternative to closed frontier APIs.
Assessment
Strengths and limitations
- Gold-medal reasoning: top-tier performance in IMO 2025, IOI 2025, and ICPC World Finals 2025, approaching frontier models with 20× fewer parameters.
- Open weights under the NVIDIA Open Model License: fully auditable and self-hostable.
- Strong agentic capabilities with native tool use, reasoning, web search, and structured JSON output on Venice.
- Dual-mode operation: explicit thinking mode for hard problems and instruct mode for fast answers.
- 256K context window and 32K max output for long-document analysis and extended generation.
- MMLU-Pro (79.8) and GPQA-Diamond (76.1) lag behind larger rivals like Qwen3.5-35B-A3B, showing knowledge gaps at scale.
- SciCode score (36.4) is moderate, indicating specialized scientific coding remains a challenge.
- Not uncensored: standard safety alignment is present, which may limit certain edge-case workflows.
- As a recent release (March 2026), production tooling and fine-tuning ecosystems are still maturing.
Specifications
Datasheet
- Maker
- NVIDIA
- Released
- March 2026
- Architecture
- Mixture-of-Experts (MoE)
- Parameters
- 30B total / 3B activated per token
- Open weights
- Yes — NVIDIA Open Model License
- Context window
- 256K tokens
- Max output
- 32.768K tokens
- Capabilities
- Function calling, Reasoning, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Mar 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia-nemotron-cascade-2-30b-a3b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.14 per 1M input tokens and $0.80 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Context window | Open weights | Price (Venice) | Best for |
|---|---|---|---|---|
| Nemotron Cascade 2 30B A3B | 256K tokens | Yes | $0.14 in · $0.80 out / 1M | Gold-medal reasoning |
| DeepSeek V3.2 | 160K tokens | Yes | $0.33 in · $0.48 out / 1M | General MoE efficiency |
| Google Gemma 4 31B Instruct | 256K tokens | Yes | $0.12 in · $0.36 out / 1M | Cost-efficient local inference |
| Kimi K2.6 | 256K tokens | Yes | $0.75 in · $3.50 out / 1M | Long-context open weights |
Compact open MoE with competition-grade math and coding performance.
Use cases
What it is good for
- 01Mathematical proof assistance and olympiad-level problem solving.
- 02Competitive programming and algorithmic code generation.
- 03Agentic workflows requiring function calling, reasoning, and web search.
- 04Long-context document analysis and structured data extraction.
- 05STEM research and education.
Prompting
Getting better results
Switch to thinking mode for complex math or coding prompts; use instruct mode for general Q&A.
Provide explicit reasoning steps in the prompt to guide the model's chain-of-thought.
Use structured output (JSON schema) for agentic tasks that require machine-readable responses.
Place key instructions at the beginning or end of long contexts for best attention.
FAQ
Frequently asked questions
Nemotron Cascade 2 30B A3B is NVIDIA's open-weight Mixture-of-Experts text model released in March 2026. It activates 3B parameters per token from 30B total, delivers gold-medal math and coding reasoning, and supports tool use, web search, and structured output.
Venice bills per token at $0.14 per 1M input tokens and $0.80 per 1M output tokens. There is no subscription required; you pay only for what you use.
Yes. The weights are released under the NVIDIA Open Model License, so you can download, inspect, and self-host the model. Venice runs it with zero retention of your prompts.
Yes. On Venice it supports function calling, reasoning, web search, and structured JSON output, making it suitable for agentic workflows.
Choose Nemotron Cascade 2 for competition-grade math and coding reasoning in a compact 30B MoE. DeepSeek V3.2 is a larger generalist MoE with cheaper output pricing, better for broad tasks where extreme reasoning specialization is less critical.
It supports up to 256K tokens of context and can generate up to 32,768 tokens in a single response.
No. It is an open-weight model but carries standard safety alignment; it is not marketed as uncensored. Venice's zero-retention privacy ensures your prompts stay private regardless.
Venice offers free trial credits for new accounts, so you can test the model without a credit card. Sustained use is billed per token.
Run Nemotron Cascade 2 30B A3B privately
No prompt logging. No data used for training.