// high-volume agentic workload
24 tickets land at once.
24 agents, one 12B-active model.
Every card is an independent Nemotron 3.5 Super VL agent loop against the same GMI endpoint: look the requester up, read service status and logs, act with 21 tools, write the resolution. Flip the thinking dial between runs and compare speed, tokens and the rubric score in the table below.
12B / 120Bactive / total params
21tools per agent
2live incidents to find
~3ksystem prompt tokens
Thinking
Context
Parallel
agents in flight at once
Tools
OpenAI function calling; falls back to JSON-in-content if the endpoint refuses
$ / M tokinoutset to your GMI price; defaults are a list-price estimate
Paste a key, pick a thinking mode, release the tickets.
0.0s
elapsed
0/24
tickets resolved
0
tokens / s, all agents
peak 0
0
output tokens
reasoning —
0
input tokens
0
tool calls
—
rubric pass
est. $0.0000
RUNS THIS SESSION
same 24 tickets, different dial positions| # | thinking | runbook | parallel | wall time | resolved | rubric | tool calls | in tok | out tok | reasoning tok | avg tok/s | peak tok/s | est. cost |
|---|