浣跨敤 Nemotron 3 Online 杞绘澗鏋勫缓鎮ㄧ殑 AI ?? 浜у搧

Nemotron 3 Online 鏄瀯寤?AI ?? 鐨勬渶浣虫ā鏉匡紝鍐呯疆 AI銆佽韩浠介獙璇併€佸叏鐞冩敮浠樸€佸崥瀹€佹枃妗ｃ€侀偖浠惰闃呫€丼EO銆佸褰╀富棰樸€佷赴瀵岀粍浠剁瓑銆?,

Nemotron 3 Live Playground

Switch between Nemotron 3 Nano and Super to compare model behavior in real time. Choose a model, ask a question, and test Nemotron 3 Super API-style prompts across long-context tasks.

缁熻

Nemotron 3 Online 鐩稿叧鐨勬暟瀛?,

Nemotron 3 Online 鍙鎮ㄥ湪鍑犲ぉ鍐呰交鏉炬瀯寤烘偍鐨?AI ??

120B / 12B

Super: 120B total parameters with 12B active

Context window up to 1M tokens

30B / 3.5B

Nano: 30B total parameters with 3.5B active

Where teams use Nemotron 3

Practical evaluation paths

These workflows are the fastest ways to validate long-context reasoning, agentic behavior, and throughput before you commit to deployment.

Long-document synthesis

Summarize reports, legal briefs, and multi-chapter research with Nemotron 3 in a single prompt window.

Multi-file code analysis

Ask the models to navigate large repos, explain architecture, and propose refactors.

Agentic tool workflows

Test multi-step planning and tool calling for research, ops, or automation tasks.

Retrieval + 1M context

Compare RAG strategies against full-context prompting to see what works best.

Local deployment planning

Evaluate quantization targets, VRAM needs, and latency before moving on-prem.

Multilingual evaluations

Run the same prompts across languages to measure consistency and quality.

Research highlights

Architecture and benchmark visuals

Figures sourced from the paper and shared assets, highlighting Nemotron 3 architecture and evaluation snapshots.

Nemotron 3 Super architecture diagram highlighting hybrid Mamba Transformer layers — Architecture overview (source: Nemotron-3 paper PDF).

Nemotron 3 Super benchmark figure summarizing accuracy and throughput tradeoffs — Benchmark snapshot (source: Nemotron-3 paper PDF).

Nemotron 3 paper figures

Paper visuals

Key diagrams and benchmark curves extracted from the paper.

Architecture and routing

MoE routing diagrams used in Nemotron 3 Nano.

The paper introduces Nemotron 3 (Nano, Super, Ultra) as a Mixture-of-Experts hybrid Mamba–Transformer family built for strong throughput and up to 1M-token context.

Most layers interleave Mamba-2 and MoE blocks with a small number of self-attention layers; larger models add LatentMoE and MTP layers for quality and faster generation.

Post-training uses multi-environment reinforcement learning to improve reasoning, multi-step tool use, and budget-controlled inference.

Nemotron 3 MoE routing diagram showing expert selection flow details — MoE routing (base)

Nemotron 3 latent MoE routing diagram for hybrid expert gating — Latent MoE routing

Accuracy-efficiency trade-off

Illustrates the accuracy-efficiency trade-off curves of Nemotron 3 Nano by varying the token budget.

Inference-time budget control lets you set a maximum token budget for the thinking trace.

When the budget is reached, appending the `</think>` token prompts the model to continue with the response based on the partial trace.

The curves below show how accuracy trades off against efficiency as the token budget changes.

Nemotron 3 AIME25 benchmark chart from the official paper snapshot — AIME25

Nemotron 3 GPQA benchmark chart from the official paper snapshot — GPQA

Nemotron 3 LiveCodeBench benchmark chart from the official paper snapshot — LiveCodeBench

Nemotron 3 MMLU Pro benchmark chart from the official paper snapshot — MMLU Pro

闆嗘垚

涓庢偍鏈€鍠滄鐨勫伐鍏烽泦鎴?,

鏃犵紳杩炴帴娴佽鐨勫钩鍙板拰鏈嶅姟锛屼互澧炲己鎮ㄧ殑宸ヤ綔娴佺▼銆?,

Google Gemini

杩欓噷鏄疓oogle Gemini鐨勬弿杩帮紝璇︾粏浠嬬粛Google Gemini鐨勫姛鑳藉拰浼樺娍銆?

浜嗚В鏇村

Replit

杩欓噷鏄疪eplit鐨勬弿杩帮紝璇︾粏浠嬬粛Replit鐨勫姛鑳藉拰浼樺娍銆?

浜嗚В鏇村

MagicUI

杩欓噷鏄疢agicUI鐨勬弿杩帮紝璇︾粏浠嬬粛MagicUI鐨勫姛鑳藉拰浼樺娍銆?

浜嗚В鏇村

VSCodium

杩欓噷鏄疺SCodium鐨勬弿杩帮紝璇︾粏浠嬬粛VSCodium鐨勫姛鑳藉拰浼樺娍銆?

浜嗚В鏇村

MediaWiki

杩欓噷鏄疢ediaWiki鐨勬弿杩帮紝璇︾粏浠嬬粛MediaWiki鐨勫姛鑳藉拰浼樺娍銆?

浜嗚В鏇村

Google PaLM

杩欓噷鏄疓oogle PaLM鐨勬弿杩帮紝璇︾粏浠嬬粛Google PaLM鐨勫姛鑳藉拰浼樺娍銆?

浜嗚В鏇村

鍔熻兘