NVIDIA: Nemotron Nano 9B V2 (free)

Capabilities

Input modalities: text

Recommended use cases

• Reasoning/planning tasks with explicit constraints and step checks
• Structured extraction or validated JSON outputs, with parser validation
• Tool-using agent tasks with evidence-grounded final answers

Skills and prompt patterns

• Mini-skill: use tools before claims, cite source/tool IDs, do not invent results, no production authority
• Reasoning guardrail: final answer must separate evidence, assumptions, and recommendation

Best practices

• Use external source evidence before final recommendation; keep local eval evidence authoritative for behavior.
• Start with low-temperature deterministic tests before creative tasks
• Log provider, returned model, latency, status, score, prompt/skill version, and redacted errors
• Persist every result immediately, including bad outcomes
• Do not test capabilities that catalog/provider metadata says are unsupported
• Prefer native tool/function calling over brittle strict JSON when tool support exists
• Validate structured outputs with an external parser; do not trust raw format claims

Test evidence

Recent records

Next test plan

researchsmokemodel_specific_optimaltools_skillsworkflow_langgraph_mockfinal_recommendation_pagevalidator_loopstructured_output_validationreasoning_planning

Research sources

External source enrichment present: yes

Recommendation status

Recommended roles

monitoring_only_until_retest

Risks and caveats

• soft: empty_output (6)
• policy: free_status_is_transient (1)