How to choose an LLM for FarkiBot

FarkiBot doesn’t pick a model for you, and it doesn’t hide which one ran. Every answer names the model that produced it, and your choice is remembered per task — so Ask, Report, and the generation tasks can each keep their own.

This page is about making that choice well.

The badges in the model picker

Badge	Meaning
Starred	Proven model-builders. The safest choice, and the right one for anything structural.
Gift-marked	Free. Exploring costs nothing.
Neither	Available, untested by us for this purpose. Fine for Ask; try it on generation and see.

The list is the live catalog from your provider, so it changes as models are released. Search it by name.

The split that matters: M1 vs M2

This is the single most useful thing to understand, and it is not obvious.

Generating an M1 model is much easier than generating an M2 metamodel, and the gap is large enough to change which model you should pay for.

	Generate M1	Generate M2
What the LLM must produce	Content that fills in a structure already given to it in the prompt	The whole structure and its notation, invented
Notation	Inherited from the metamodel	Must be authored — including inline SVG icons inside JSON strings, where one mis-escape invalidates the whole file
Validation	Against the importer and the governing metamodel	Against the importer
Free models	Often succeed	Usually fail

In our own testing, nvidia/nemotron-3-ultra-550b:free produced a good Business Model Canvas M1 at no cost — and failed three attempts at the M2 for the same subject.

That asymmetry is not a coincidence, and it leads to a strategy:

Generate the metamodel once with a starred model, then build models on it for free or nearly free.

You pay only where the stakes and the difficulty actually are. A metamodel is authored once; models are authored constantly.

Task by task

Task	What to use	Why
Ask	Anything, including free models	Reading a model and answering from it is well within a mid-tier model. Try a free one first.
Report	Mid-tier or better	Document quality — structure, judgement, restraint — is where stronger models show. A weak model produces something that reads like a report but says nothing.
Generate M1	Free or open-weight workhorses	The metamodel constrains it. Kimi-class models are excellent here.
Generate M2	Starred models only	A flaw in the metamodel echoes through every model later built on it.
Modify M1	Starred models	The most demanding task — see below.

Why Modify M1 is the hardest task

Modifying a model is harder than generating one from scratch, which surprises most people.

FarkiBot rewrites the entire model with your change applied. So the LLM’s real job isn’t the change you asked for — it’s faithfully reproducing everything you didn’t ask about, at length, without dropping anything. That is a preservation task, and weak models are bad at it.

The diff will show you if content went missing. Use a strong model and you’ll rarely need it to.

What it costs

Real numbers from our own testing, so you have a sense of scale:


A valid M1 from Kimi K3, first attempt	About $0.16 — roughly a tenth of what a frontier model charges for the same job, and faster
A Business Model Canvas M1 from a free model	$0.00
A researched answer vs a plain one	Roughly 3–4× more — the search results are added to the prompt

Most of what you pay for is context: the model export that goes into the prompt. So the cheapest lever you have isn’t the model, it’s the Context chip. Narrowing a question to the two packages it’s actually about will save you more than switching models.

Cost visibility depends on the provider. OpenRouter reports the cost of each run and FarkiBot shows it. Anthropic and Google report token counts but no price, so those runs show “cost not reported” — check your provider console for spend.

Getting to zero cost

If you want to explore FarkiBot without spending anything:

Create a free OpenRouter account
Use gift-marked (:free) models
Stick to Ask and Generate M1
Import a metamodel from the Farketplace rather than generating one

Free models are rate-limited to 50 requests per day. Buying about $10 of credits once raises that to 1,000 per day — and the free models remain free.

When a model isn’t up to the task

FarkiBot tells you honestly rather than importing something broken. A generation that fails validation after three attempts produces a failure card that explains why, offers the raw output for download, and lets you Try again with a different model.

Some specific failures and what they mean:

Message	What it’s telling you
”Generation failed validation”	The model couldn’t produce a valid model file. Retry with a starred model.
”…hidden reasoning” errors	A reasoning model burned its output budget thinking. FarkiBot already tried raising the limit and disabling reasoning — this model isn’t going to work here.
”…cut off at the output token limit”	The file didn’t fit. FarkiBot raises the limit automatically up to the model’s maximum; if it keeps happening, narrow the brief or pick a model with a larger output cap.
”…remaining credit can’t cover the full output budget”	Your key’s spend cap can’t pre-authorise the full output. Raise the cap or top up for large generations.

See FarkiBot Reference for the full list.

A practical starting recipe

OpenRouter key with a $10 credit limit
Ask — start with a free model, upgrade only if answers feel thin
Generate M2 — a starred model, once
Generate M1 — Kimi-class or free, repeatedly
Modify M1 — a starred model, and read the diff