Choose and check a model

Verbiqo needs a model to explain what it reads and to write. You do not run one yourself unless you want to. Choose where yours comes from, enter it in Settings, then check it once before relying on it.

Where your model comes from

Your organization’s endpointA model server or gateway your organization already runs. The usual choice at work. An approved cloud APIThe Anthropic API, or an OpenAI-compatible service your organization allows. A model on this computerOllama on your Mac. Private, but no laptop model has passed the check yet.

Your organization’s endpoint

Ask your administrator for: the endpoint address, the model name, a key if it needs one, and the CA certificate if the server uses your organization’s own.
Enter them: menu bar icon › Settings… › AI provider, usually OpenAI-compatible. Press Test & Load Models, choose the Model, then Save. A private CA goes under Settings › Trusted CA.
Your administrator may have set this already for everyone, in which case the fields are filled in and locked.

Running that server is covered in Run a model server for your team.

An approved cloud API

Only with your organization’s approval. Your question and what Verbiqo reads for it go to that provider. Verbiqo asks you to confirm the endpoint before anything is sent.
Enter it: Settings… › AI provider, then Anthropic API or OpenAI-compatible, the key, and the model. The key is kept in the Keychain or Credential Manager.
Not yet on the tested list. Run the model check before relying on it.

A model on this computer

Not yet verified. The models that passed ran on GPU servers. The only small model checked on a Mac, qwen2.5:1.5b, did not pass: it missed the cause in an investigation and changed values in rewrites.
What it takes: a model of the size that passed, about 30 billion parameters at 4-bit, needs about 18 GB for the model alone. A Mac with plenty of memory may run one, but that has not been checked.
To try it: install Ollama, pull a model, choose Ollama in Settings… › AI provider, and set Context window as in Ollama. Then run the model check, and rely on it only if it passes.

What a model needs

Settings: Context window 32768, Timeout (ms) 300000, Temperature 0.2 unless the server sets it. See Requirements and settings.
An instruct model, trained to answer directly. Models that reason at length work, but each step takes minutes: see Reasoning models.
Tool calling makes database questions faster. Without it Verbiqo works by plain text.

Two set-ups have passed and answer real questions well, including root-cause investigations of a MySQL cluster outage (October 2026): Qwen3-30B-A3B-Instruct-2507 on llama.cpp and Qwen3.6-35B-A3B on vLLM. Their results are under Tested models, and their server settings are in Run a model server for your team.

Before you start: check the model

The model check runs Verbiqo’s own steps against your model and says whether it is fit, how long each step takes, and what to change if it is not. It uses built-in test data: it never connects to Prometheus, a database or a file. Only the model is called.

Point it at your model
With no options it uses the model set in Verbiqo’s Settings. To check another, give its address and the name the server serves.

Run it

macOS, with the command line tool installed:

verbiqo model-check | tee model-check.md

Or without installing it:

/Applications/Verbiqo.app/Contents/Helpers/verbiqo model-check | tee model-check.md

Linux, from the folder you unpacked the command line tool into:

./verbiqo model-check --provider openai \
  --endpoint http://127.0.0.1:8000/v1 \
  --model "my-model" --timeout-ms 300000 | tee model-check.md

Windows, from the cli folder of your Verbiqo install:

verbiqo.exe model-check > model-check.md

Read the result
It takes a minute or two on a fast model. The report says Fit for Verbiqo or Not fit as configured, with a line per check and a list of what to change.

If your server uses a certificate from your organization’s own authority, the check reports the model as unavailable until it trusts that authority: add the authority to this computer’s trust store or choose it under Settings › Trusted CA; the terminal trusts both by itself. On a terminal without the app, export NODE_EXTRA_CA_CERTS=/path/to/your-ca.crt also works (on Windows, set NODE_EXTRA_CA_CERTS=C:\path\to\your-ca.crt). If the server needs a key, set VERBIQO_API_KEY or let it read the one saved in Settings.

What is checked

Check	What it runs	Passes when
Prometheus investigation	A known MySQL outage on built-in data, through Verbiqo’s steps: reading the question, choosing what to measure, writing the answer.	Every step answers, and the model’s own text names the host that was down, the outage, and what led up to it, with no figure that was never measured.
Database statement	The first step of a database question, where the model chooses a read-only statement.	It produces a statement, by tool call or by plain text. The report says which: tool calls are faster.
Rewrites keep every value	12 technical sentences in three modes, with versions, hosts, sizes, times and flags.	At least 9 in 10 come back with every value intact.

Every model call is timed and inspected: whether the model reasoned before answering, repeated itself, came back empty, or failed. A pass on this fixed check is a baseline, not a guarantee: a model that fails it will not do well on real questions.

Tested models

Historical results of the model check, with the hardware and settings recorded at the time. The exact server build versions were not recorded. The service units in the model-server guide were written later and have not been runtime-validated on a GPU server. None of these checks ran on a laptop. Run the check on yours: a different quantization, server version or setting can change the result.

Model	Run on	Result	Notes
Qwen3.6-35B-A3B (AWQ), served as `qwen3-30b-instruct`	vLLM, the historically checked configuration (deployment recipe): 32,768-token context, reasoning off, MTP speculative decoding, tool calling on	Fit on the model check (1 Oct 2026)	The fastest checked: investigation in 5 seconds (steps of 1.1 to 2.4 s), median call 0.4 s. Statements chosen by tool call. 35 of 36 rewrites kept every value; the one refused moved a size away from the word it belonged to. Never reasoned, repeated or came back empty.
Qwen3-30B-A3B-Instruct-2507 (Q4_K_M), served as `qwen3-30b`	llama.cpp, the historically checked configuration (deployment recipe): 32,768-token context, two GPUs	Fit on the model check (1 Oct 2026)	Investigation in 24 seconds (answer step 16 s), median call 0.8 s. Statements chosen by tool call. All 36 rewrites kept every value. Never reasoned, repeated or came back empty.
qwen3.8-27b (int4)	vLLM, one A100 40 GB, 32,768-token context, tool calling on, behind a LiteLLM gateway	Fit on the model check (1 Oct 2026)	About a minute for the check; 3 to 11 seconds per step. Tools used. 35 of 36 rewrites kept every value. A hybrid reasoning model: Verbiqo turns its reasoning off per request. On a real MySQL outage it chose mostly host and disk metrics over MySQL’s own; see Instruct models.
qwen2.5:1.5b	Ollama on a Mac	Not fit (1 Oct 2026)	Fast, but its investigation answer did not name the host that was down, it did not produce a statement, and it changed values in several rewrites. Too small for Verbiqo’s work.

Seen in use, not yet checked

Model	What was seen	What to do
Mistral-Small-24B-Instruct-2501 (AWQ)	On a long investigation it repeated one sentence until its output limit. Verbiqo now cuts a repeated answer and keeps its own findings, but the explanation is lost.	Add a small repetition penalty on the server, or use Mistral Small 3.2, which Mistral says repeats less. Run the model check either way.
Mistral-Small-3.2-24B-Instruct-2506	A candidate for a single 40 GB GPU, not yet checked.	Run the model check and compare with the results above.
Qwen3-30B-A3B-Instruct-2507 (FP8)	On vLLM, one A100 40 GB: a real MySQL outage, with good metric choices. Without a temperature set, its explanation of the same outage varied between runs.	Set Temperature to 0.2, and run the model check.

Server notes

Server settings for llama.cpp and vLLM, with a service account, a key and TLS, are in Run a model server for your team.

Behind a gateway

A gateway such as LiteLLM can serve a model under another name. The model check reports the model the server says answered, and warns when it differs from the one asked for: test, and record, the model that actually answers. A gateway also needs its own timeout and context size set: an example LiteLLM entry.

Ollama

Ollama sizes each model’s context window itself. Set Context window in Settings to the size ollama ps shows for the model, or Verbiqo assumes a small window and trims what it reads to fit.

Instruct models first, and still check them

Verbiqo’s model steps are short, structured choices: what to measure, which statement to run, and a short explanation of findings Verbiqo has already measured. Instruct models, trained to answer directly, suit them best. Being an instruct model is not enough on its own: Mistral-Small-24B-Instruct-2501 is one, and it repeated itself on long evidence. The model check is what tells. A hybrid reasoning model with its reasoning off works in its weaker mode, and with its reasoning on it is slow. On the same MySQL outage, an instruct model (Qwen3-30B-A3B-Instruct-2507) chose MySQL’s own metrics, while a hybrid model with reasoning off chose mostly host and disk metrics and read the quiet disks as the cause.

Reasoning models

Some models reason at length before they answer. Verbiqo’s model steps are short, structured choices over what Verbiqo has already measured, and long reasoning only makes each step slower: minutes instead of seconds.

Qwen3: Verbiqo asks it not to reason, in each request. The model on the server is unchanged, so other applications sharing it still get its reasoning. VERBIQO_MODEL_THINKING=on lets it reason in Verbiqo too.
Models that always reason (Qwen3 “Thinking” releases and similar): Verbiqo gives a step that ran out of room one more try with more, and shows its own findings if the answer still is not written. Expect each step to take longer, and raise Timeout in Settings.

When a model sends back reasoning, Verbiqo says so once in the conversation: it works, but each step takes about ten times longer than with an instruct model.

Whatever the model, every figure in an answer comes from what Verbiqo measured, and a figure the model invented is marked. A weaker model gives a weaker explanation, not wrong numbers.