{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b","slug":"nvidia-nemotron-3-super-120b-a12b","name":"NVIDIA: Nemotron 3 Super","type":"model","url":"https://openrouter.ai/models/nvidia~nemotron-3-super-120b-a12b","page_url":"https://unfragile.ai/nvidia-nemotron-3-super-120b-a12b","categories":["ai-agents"],"tags":["nvidia","api-access","text"],"pricing":{"model":"paid","free":false,"starting_price":"$9.00e-8 per prompt token"},"status":"active","verified":false},"capabilities":[{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_0","uri":"capability://text.generation.language.sparse.mixture.of.experts.inference.with.dynamic.parameter.activation","name":"sparse-mixture-of-experts inference with dynamic parameter activation","description":"Nemotron 3 Super uses a hybrid Mamba-Transformer architecture with sparse Mixture of Experts (MoE) routing that activates only 12B of 120B parameters per forward pass. The model employs learned gating mechanisms to route tokens to specialized expert sub-networks, reducing computational cost while maintaining model capacity. This sparse activation pattern is computed dynamically based on input tokens, enabling efficient inference on consumer-grade hardware without quantization.","intents":["Run a 120B-parameter model on limited compute budgets without sacrificing reasoning quality","Deploy multi-agent systems where inference latency and throughput matter more than raw parameter count","Reduce inference costs when operating at scale by activating only 10% of model parameters per token"],"best_for":["Teams building cost-sensitive multi-agent applications requiring complex reasoning","Researchers benchmarking sparse MoE architectures against dense models","Production systems where inference throughput and latency are critical constraints"],"limitations":["MoE routing adds ~50-100ms latency overhead per inference step compared to dense models due to expert selection computation","Expert load balancing can be uneven across tokens, potentially causing some experts to be underutilized","Sparse activation pattern is non-deterministic during training; inference behavior may differ slightly across runs without fixed random seeds","Memory footprint remains ~120B parameters even though only 12B activate, requiring sufficient VRAM for model loading"],"requires":["API access via OpenRouter or compatible inference endpoint","Minimum 40GB VRAM for single-GPU inference (or distributed inference setup)","Network connectivity for API calls; no local deployment option documented"],"input_types":["text (natural language prompts)","code (for code understanding and generation tasks)","structured prompts with system instructions"],"output_types":["text (natural language responses)","code (generation, explanation, refactoring)","structured reasoning traces (chain-of-thought)"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_1","uri":"capability://planning.reasoning.multi.agent.conversation.orchestration.with.long.context.reasoning","name":"multi-agent conversation orchestration with long-context reasoning","description":"Nemotron 3 Super is optimized for multi-agent applications where multiple specialized agents coordinate to solve complex tasks. The model maintains coherent context across extended conversations, tracking agent roles, responsibilities, and shared state. The architecture supports deep reasoning chains where agents build on each other's outputs, with the sparse MoE design ensuring each agent's specialized reasoning path activates relevant experts without full model overhead.","intents":["Coordinate multiple AI agents (planner, executor, validator) in a single conversation without context loss","Build systems where agents reason about each other's outputs and refine solutions iteratively","Maintain consistent agent personas and role-based reasoning across 10K+ token conversations"],"best_for":["Teams building complex agentic workflows (research, planning, code generation with review)","Applications requiring multi-turn reasoning where agents depend on prior agent outputs","Systems where agent specialization matters (e.g., one agent for planning, one for execution, one for validation)"],"limitations":["No explicit agent state management built-in; requires external orchestration layer to track agent roles and conversation history","Context window size not explicitly documented; typical for 120B models is 4K-8K tokens, limiting very long multi-agent chains","Agent coordination logic must be implemented in prompting/system instructions; no native multi-agent protocol support","No built-in memory persistence across sessions; each conversation starts fresh unless explicitly managed"],"requires":["API access via OpenRouter or compatible endpoint","External orchestration framework (e.g., LangChain, AutoGen, custom agent loop)","Structured prompting with clear agent role definitions in system instructions"],"input_types":["text (agent instructions, conversation history, task descriptions)","structured agent messages (role, content, metadata)"],"output_types":["text (agent responses, reasoning traces)","structured agent actions (tool calls, decisions, state updates)"],"categories":["planning-reasoning","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_2","uri":"capability://code.generation.editing.code.generation.and.multi.file.refactoring.with.context.awareness","name":"code generation and multi-file refactoring with context awareness","description":"Nemotron 3 Super generates code across multiple programming languages and can understand multi-file codebases for refactoring tasks. The model uses its extended context window and reasoning capabilities to track dependencies between files, suggest structural improvements, and generate coherent changes across a codebase. The sparse MoE architecture allows code-specific experts to activate for syntax-aware generation while general reasoning experts handle architectural decisions.","intents":["Generate production-quality code in Python, JavaScript, Go, Rust, and other languages from natural language specifications","Refactor multi-file codebases by understanding cross-file dependencies and suggesting consistent changes","Explain complex code logic and suggest optimizations based on architectural patterns"],"best_for":["Solo developers and small teams building full-stack applications","Teams migrating codebases between languages or frameworks","Development teams using AI-assisted code review and refactoring workflows"],"limitations":["No built-in AST parsing or syntax validation; generated code requires testing and linting","Context window limits multi-file refactoring to codebases under ~4K-8K tokens; very large projects require chunking","No IDE integration documented; requires manual copy-paste or API integration for seamless workflow","Code generation quality varies by language; better performance on Python/JavaScript than niche languages"],"requires":["API access via OpenRouter","Code context provided as text (no direct file system access)","Testing/linting tools to validate generated code before deployment"],"input_types":["text (code snippets, natural language specifications)","code (existing codebase for refactoring context)"],"output_types":["code (generated functions, classes, full files)","text (explanations, refactoring suggestions)"],"categories":["code-generation-editing","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_3","uri":"capability://planning.reasoning.complex.reasoning.with.chain.of.thought.decomposition","name":"complex reasoning with chain-of-thought decomposition","description":"Nemotron 3 Super excels at breaking down complex problems into reasoning steps, generating explicit intermediate reasoning before final answers. The model can produce detailed chain-of-thought traces for mathematical problems, logical reasoning, and multi-step planning tasks. The hybrid Mamba-Transformer architecture provides both efficient sequence modeling (Mamba) and attention-based reasoning (Transformer), enabling coherent multi-step reasoning without excessive parameter activation.","intents":["Solve complex math problems by generating step-by-step reasoning before final answers","Decompose ambiguous user requests into structured reasoning steps and sub-tasks","Generate transparent decision-making traces for auditable AI systems"],"best_for":["Educational applications requiring explainable reasoning","Compliance-sensitive systems where decision transparency is required","Research teams benchmarking reasoning capabilities of sparse models"],"limitations":["Chain-of-thought generation increases latency by 2-5x compared to direct answers","Reasoning quality degrades on very specialized domains (e.g., advanced physics) without domain-specific fine-tuning","No built-in validation of reasoning correctness; intermediate steps may contain logical errors","Reasoning traces are generated text, not structured data; parsing requires additional NLP"],"requires":["API access via OpenRouter","Prompting strategy that explicitly requests step-by-step reasoning (e.g., 'Think step by step')"],"input_types":["text (problem statements, questions, ambiguous requests)"],"output_types":["text (reasoning traces, step-by-step solutions, final answers)"],"categories":["planning-reasoning","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_4","uri":"capability://tool.use.integration.api.based.inference.with.streaming.and.batch.processing","name":"api-based inference with streaming and batch processing","description":"Nemotron 3 Super is accessed exclusively through OpenRouter's API, supporting both streaming (token-by-token) and batch inference modes. The API abstracts away the underlying sparse MoE complexity, presenting a standard LLM interface. Streaming enables real-time response generation for interactive applications, while batch processing allows cost-optimized throughput for non-latency-sensitive workloads. The sparse activation is handled transparently by the inference backend.","intents":["Stream model responses in real-time for interactive chatbots and web applications","Process large batches of prompts asynchronously for cost-optimized throughput","Integrate Nemotron 3 Super into existing LLM applications via standard OpenAI-compatible API"],"best_for":["Web applications requiring real-time streaming responses","Batch processing pipelines for content generation, summarization, or classification","Teams already using OpenRouter for multi-model inference"],"limitations":["No local deployment option; all inference requires network round-trips to OpenRouter","API rate limits and quota management required for high-volume applications","Streaming adds overhead compared to batch; per-token latency is higher than batch processing","No fine-tuning support documented; model weights are fixed"],"requires":["OpenRouter API key","Network connectivity","HTTP client library (Python requests, JavaScript fetch, etc.)","Paid account with sufficient credits for inference costs"],"input_types":["text (prompts, messages)"],"output_types":["text (streaming tokens or complete responses)","structured metadata (token counts, finish reasons)"],"categories":["tool-use-integration","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_5","uri":"capability://text.generation.language.knowledge.synthesis.and.summarization.from.long.documents","name":"knowledge synthesis and summarization from long documents","description":"Nemotron 3 Super can process and synthesize information from extended documents, generating summaries, extracting key points, and answering questions about document content. The model's extended context window and efficient sparse activation enable processing of longer documents than typical dense models without excessive latency. The reasoning capabilities allow nuanced synthesis rather than simple extractive summarization.","intents":["Summarize long research papers, articles, or reports into concise key points","Answer specific questions about document content without requiring full document re-reading","Extract structured information (entities, relationships, key findings) from unstructured documents"],"best_for":["Research teams processing large volumes of academic papers","Content platforms requiring automated summarization","Knowledge workers synthesizing information from multiple sources"],"limitations":["Context window limits document length to ~4K-8K tokens; very long documents require chunking and multi-pass processing","Summarization quality depends on document structure; poorly formatted documents may produce incoherent summaries","No built-in document parsing; requires pre-processing to extract text from PDFs, images, or other formats","Extractive accuracy varies; model may hallucinate details not present in source documents"],"requires":["API access via OpenRouter","Document text pre-processed and provided as input (no direct file access)","Document length under context window limit"],"input_types":["text (document content, questions about documents)"],"output_types":["text (summaries, extracted information, answers)"],"categories":["text-generation-language","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"openrouter-nvidia-nemotron-3-super-120b-a12b__cap_6","uri":"capability://text.generation.language.instruction.following.and.task.specific.adaptation","name":"instruction-following and task-specific adaptation","description":"Nemotron 3 Super is trained to follow detailed instructions and adapt behavior based on system prompts and task specifications. The model can adjust tone, style, output format, and reasoning approach based on explicit instructions. This capability enables single-model deployment across diverse applications without model switching. The sparse MoE design allows task-specific experts to activate based on instruction content, improving efficiency for specialized tasks.","intents":["Configure model behavior for specific applications (e.g., customer support, technical writing, creative content) via system prompts","Request specific output formats (JSON, markdown, code blocks) and have the model reliably produce them","Adapt model personality and communication style for different user personas or use cases"],"best_for":["Multi-purpose applications requiring diverse model behaviors from single deployment","Teams building customizable AI assistants for different user segments","Applications where output format consistency is critical (e.g., structured data extraction)"],"limitations":["Instruction-following quality degrades with conflicting or ambiguous instructions","No guarantee that model will respect all instructions; complex constraints may be violated","System prompt injection attacks possible if user input is not properly sanitized","Output format compliance requires careful prompt engineering; no built-in schema validation"],"requires":["API access via OpenRouter","Well-crafted system prompts and instructions","Input validation to prevent prompt injection"],"input_types":["text (system prompts, user instructions, task descriptions)"],"output_types":["text (responses in requested format, structured data)"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":23,"verified":false,"data_access_risk":"high","permissions":["API access via OpenRouter or compatible inference endpoint","Minimum 40GB VRAM for single-GPU inference (or distributed inference setup)","Network connectivity for API calls; no local deployment option documented","API access via OpenRouter or compatible endpoint","External orchestration framework (e.g., LangChain, AutoGen, custom agent loop)","Structured prompting with clear agent role definitions in system instructions","API access via OpenRouter","Code context provided as text (no direct file system access)","Testing/linting tools to validate generated code before deployment","Prompting strategy that explicitly requests step-by-step reasoning (e.g., 'Think step by step')"],"failure_modes":["MoE routing adds ~50-100ms latency overhead per inference step compared to dense models due to expert selection computation","Expert load balancing can be uneven across tokens, potentially causing some experts to be underutilized","Sparse activation pattern is non-deterministic during training; inference behavior may differ slightly across runs without fixed random seeds","Memory footprint remains ~120B parameters even though only 12B activate, requiring sufficient VRAM for model loading","No explicit agent state management built-in; requires external orchestration layer to track agent roles and conversation history","Context window size not explicitly documented; typical for 120B models is 4K-8K tokens, limiting very long multi-agent chains","Agent coordination logic must be implemented in prompting/system instructions; no native multi-agent protocol support","No built-in memory persistence across sessions; each conversation starts fresh unless explicitly managed","No built-in AST parsing or syntax validation; generated code requires testing and linting","Context window limits multi-file refactoring to codebases under ~4K-8K tokens; very large projects require chunking","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.05,"quality":0.39,"ecosystem":0.24,"match_graph":0.25,"freshness":0.75,"weights":{"adoption":0.35,"quality":0.2,"ecosystem":0.1,"match_graph":0.3,"freshness":0.05}},"observed_outcomes":{"matches":0,"success_rate":0,"avg_confidence":0,"top_intents":[],"last_matched_at":null},"maintenance":{"status":"active","updated_at":"2026-05-24T12:16:24.484Z","last_scraped_at":"2026-05-03T15:20:45.776Z","last_commit":null},"community":{"stars":null,"forks":null,"weekly_downloads":null,"model_downloads":null,"model_likes":null}},"distribution":{"claim_url":"https://unfragile.ai/submit?claim=nvidia-nemotron-3-super-120b-a12b","compare_url":"https://unfragile.ai/compare?artifact=nvidia-nemotron-3-super-120b-a12b"}},"signature":"5J9C8+2u/X36GVax9ToYi5JEdIGkWfuFkig1uiOkrIuWhs19bSEGtkzAfFrdKMrzN5mP3Cv7+CVqZPwkqDCaCg==","signedAt":"2026-06-22T11:24:57.841Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/nvidia-nemotron-3-super-120b-a12b","artifact":"https://unfragile.ai/nvidia-nemotron-3-super-120b-a12b","verify":"https://unfragile.ai/api/v1/verify?slug=nvidia-nemotron-3-super-120b-a12b","publicKey":"https://unfragile.ai/api/v1/trust-passport-public-key","spec":"https://unfragile.ai/trust","schema":"https://unfragile.ai/schema.json","docs":"https://unfragile.ai/docs"}}