# Models

Change free token limits and understand which model and engine settings the seller can serve.

Source: https://docs.gyblos.com/en/sellers/models-offers

The default installation serves pinned **Qwen3 0.6B through local Ollama**, with reasoning `off`.
This is the model validated end to end for this onboarding. A catalog entry or valid TOML file does not
mean that every model has been tested.

## Free limits

Edit the existing `[[offers]]` entry in `~/gyblos-seller/gyblos.toml`. The following complete example
halves the total daily limit and lowers the per-buyer limit. It deliberately leaves publication off;
if already listed, preserve your existing `[discovery]` setting when editing just the caps.

```toml
version = 1

[node]
network = "devnet"

[discovery]
publish = false

[[offers]]
model = "qwen3:0.6b"
reasoning = ["off"]
engine = { kind = "ollama", url = "http://127.0.0.1:11434/" }
free = { daily_token_cap = 100000, per_buyer_cap = 2000 }
```

Both caps are positive integers, with the per-buyer cap no higher than the daily total. They count
**input plus output** per UTC day. Admission reserves input plus the requested maximum output, so a
request can be refused before the nominal remaining balance reaches zero. Valid final usage releases
the unused reservation.

The counters are in memory and reset on restart. Per-buyer limits are best effort; they are not
persistent billing or a guarantee against someone creating another buyer identity.

### Apply changes

```sh
cd ~/gyblos-seller
docker compose run --rm seller gyblos config validate --config /config/gyblos.toml
```

After `OK`, apply the edit:

```sh
docker compose up -d --force-recreate seller
docker compose exec seller gyblos seller status
```

Check `RUNNING` and the offered model. Use `seller status --online` if publication is enabled.
If it fails, use the reported code in [troubleshooting](/en/help/troubleshooting). Configuration is read
at startup; editing the file alone does not change the running seller.

## Models and engines

Changing `model` does not download a model or start an engine. The installer downloads only
`qwen3:0.6b`. Ollama startup currently accepts the pinned Qwen3 0.6B, a literal loopback root URL,
no API-key variable and reasoning `off`. Keep its Docker URL at `http://127.0.0.1:11434/`.
The two containers share a network namespace; a host service or another container has different addressing.

OpenAI-compatible and Anthropic-compatible plugins also exist. They require explicit context size,
reasoning `off` and a compatible model-list endpoint as well as generation. Anthropic also requires an
API-key environment variable and an API root without `/v1`. These are advanced paths, outside the
local installation tutorial. See the [configuration reference](/en/reference/configuration).

> **A free offer can still cost you money**
>
>
> A hosted model provider can charge the seller for requests. Keep provider
> credentials in environment variables, never in TOML. Do not enable a remote
> provider until you understand its costs and data policy.

This tutorial provisions a free-only seller role. Paid offers require separate payment-role setup;
adding a `price` table alone does not complete that setup. Paid devnet operation remains outside this onboarding.

**Next:** [Operate and update](/en/sellers/operate).