Models in Agent mode

Which models drive an agent run, measured rather than assumed.

Agent mode asks a model to do something a chat model is never asked to do: call a tool, read what comes back, call another, and finish by recording claims that cite what it read. A model that writes beautiful prose about a database and never calls a tool answers nothing here. So the question these pages answer is not what a model knows but what it DOES on a run, and every figure comes from a run whose ledger is on disk.

Thirty models are supported. Each cleared all six surfaces — Investigate, Optimize, Assess, Operate, Analyze and Plan — five consecutive times: 900 runs, 900 passes.

Twenty-two of the thirty cleared them at the 90-second per-turn limit the product ships. Eight carry a 150-second limit of their own — qwen3.5:9b, gemma4:12b, gemma4:31b, nemotron-3.5-lightning:30b, nemotron-3-nano:30b, muse-glimmer:latest and qwen3.6:27b — because one turn of theirs does not fit inside 90 while every other surface does. The limit stays 90 for every other model, and each page says what its model needed.

Model	Served through	Disk	Median run	Slowest run
qwen2.5:7b	Ollama	4.7 GB	6s	21s
gemini-3.5-flash-lite	Gemini API	—	10s	21s
gemma4:12b	Ollama	7.6 GB	11s	63s
granite4.1:8b	Ollama	5.3 GB	11s	22s
qwen3.5:4b	Ollama	3.4 GB	11s	167s
cogito:14b	Ollama	9.0 GB	11s	20s
ministral-3:8b	Ollama	6.0 GB	13s	26s
ministral-3:14b	Ollama	9.1 GB	16s	56s
qwen2.5:14b	Ollama	9.0 GB	18s	65s
qwen2.5-coder:14b	Ollama	9.0 GB	20s	37s
qwen2.5:32b	Ollama	19 GB	20s	36s
cogito:32b	Ollama	19 GB	23s	39s
granite4.1:30b	Ollama	17 GB	26s	52s
nemotron3:33b	Ollama	27 GB	21s	119s
nemotron-3.5-lightning:30b	Ollama	25 GB	29s	145s
ornith:9b	Ollama	5.5 GB	31s	118s
qwen3.5:9b	Ollama	5.8 GB	33s	98s
gemma4:26b	Ollama	16 GB	36s	92s
ornith:35b	Ollama	21 GB	36s	150s
qwen3:8b	Ollama	5.2 GB	36s	108s
qwen3:14b	Ollama	9.3 GB	41s	303s
granite4.2:8b	Ollama	5.3 GB	46s	116s
nemotron-3-nano:30b	Ollama	24 GB	46s	190s
qwen3:4b	Ollama	2.5 GB	72s	160s
qwen3.8:latest	Ollama	19 GB	72s	195s
qwen3.6:27b	Ollama	17 GB	82s	252s
gemma4:31b	Ollama	19 GB	107s	398s
muse-glimmer:latest	Ollama	18 GB	115s	285s

The durations are from one machine and are comparable with each other rather than portable: every figure was taken the same way, on the same database, through the same six surfaces. What they are for is choosing between these thirty — the fastest reaches the same verdicts as the slowest in an eighteenth of the time.

One page per model version. Sizes are rows inside it, because ollama pull qwen3:4b is how a size is chosen and because the interesting fact is usually the difference between two sizes of the same model.

Configuring one — local or hosted	setup.md
How these numbers were produced, and where they stop being safe	methodology.md
Measuring a model these pages do not cover	testing-your-own.md
Giving a model settings dbportal has never measured	model-tuning.md

What was measured

Six surfaces, one question each, against the embedded SQLite sample:

Surface	Question	What it takes to pass
Investigate	“What tables are in this database and how do they relate?”	a report, cited
Optimize	“Why is the employee listing query slow?”	a plan comparison or a grounded index recommendation, then a report
Assess	“Where is this database’s data incomplete or surprising?”	a table profiled, then a report
Operate	“What is currently happening on this database?”	readings, then a report
Analyze	“Which part of the company costs us the most in salary?”	a read, an answer PRESENTED, then a report
Plan	“What tables are in this database and how do they relate?”	a runnable statement, or an explicit refusal

Five consecutive passes is what makes a surface count, and the bar is consecutive on purpose: a model that answers four times in five is not one you would put in front of a user, and a single pass says nothing at all about the fifth.

Choosing between them

If you want	Take	Why
the fastest local model	qwen2.5:7b	6s median, 4.7 GB, and nothing it did on thirty runs took longer than 21s
the smallest download	qwen3:4b	2.5 GB — and not the fastest: size buys memory, not speed
the steadiest	cogito:14b	20s is the shortest slowest-run on the roster, and under twice its own median
no local hardware at all	gemini-3.5-flash-lite	the one hosted model, 10s median, and it needs a key

What is not here

Models that do not clear all six surfaces are not listed and not supported. Several came close — one cell short, usually the same one — and being close is not the claim this product makes. Measuring one yourself is testing-your-own.md.