▤
Security questionnaire autofill
ComplianceAuto-answer a prospect's security questionnaire from your own policy library — grounded, with citations and a 'needs human review' flag.
The routing recipe
Job typeRAG / grounded Q&APolicyBalancedConstraintsctx ≥ 128k
routes to
NVIDIA Nemotron 3 Super 120B-A12B nvidia
nvidia:nvidia/nemotron-3-super-120b-a12b
89.1
match score
90.1
task fit
100
cost
70
speed
Why this model from benchmarks & capabilities
- "RAG / grounded Q&A" leans hardest on instruction; NVIDIA Nemotron 3 Super 120B-A12B scores 85/100 there — #14 of 40.
- Strength on this class of work shows up on MMLU-Pro, DROP, Natural Questions, where NVIDIA models are competitive.
- Holds the 32k+ context this job needs (this model: 262k).
- Cost: runs on NVIDIA free credits / self-hostable NIM — effectively $0 marginal. Speed: ~150 tok/s (estimated).
- Chosen over Llama 4 Maverick 17B-128E Instruct for faster.
| Capability the job needs | Weight | NVIDIA Nemotron 3 Super 120B-A12B | Catalog rank |
|---|---|---|---|
| instruction | 0.90 | #14 of 40 | |
| long_context | 0.90 | #1 of 40 | |
| knowledge | 0.80 | #10 of 40 |
Relevant benchmarksMMLU-Pro DROP Natural Questions RULER (long-context)
Why it qualifiesHolds the 32k+ context this job needs (this model: 262k).
EconomicsCost: runs on NVIDIA free credits / self-hostable NIM — effectively $0 marginal.
Speed: ~150 tok/s (estimated). vs runner-upChosen over Llama 4 Maverick 17B-128E Instruct for faster. (Llama 4 Maverick 17B-128E Instruct, score 87.7)
Speed: ~150 tok/s (estimated). vs runner-upChosen over Llama 4 Maverick 17B-128E Instruct for faster. (Llama 4 Maverick 17B-128E Instruct, score 87.7)
Example result
Sample input ▾
Question: 'Do you encrypt customer data in transit and at rest?' Context: 'Policy ENC-2: All data encrypted at rest with AES-256 (KMS). TLS 1.2+ enforced on all public endpoints.'
Answer: Yes. Customer data is encrypted at rest with AES-256 (managed via KMS) and in transit via TLS 1.2+ on all public endpoints.
Source: Policy ENC-2.
needs_review: false
Test it on your own data