Choosing a model
When to leave the model on Auto, when to force one, and what Auto actually does with your message.
Models
Auto is the default, and it picks a model for each message you send.
Most of the time, leaving it alone is the right call. Three situations justify forcing a model.
When to force a model
- A hard code task you want handled once. Force Kimi K3, the strongest of the three.
- A long build that stays on one job. Force GLM 5.2, described as the long horizon coding model with a 1M context.
- A lot of straightforward writing at the lowest price. Force MiniMax M3.
Outside those three, Auto already lands on the same answer more often than not.
What Auto actually does
A small classifier reads the first 2 000 characters of your message. It sorts the message into one of four buckets, then applies this ladder.
| Bucket | What it means | Model |
|---|---|---|
| question | You are asking, not ordering. | MiniMax M3 |
| simple | A cosmetic change in one place, one file at most. | MiniMax M3 |
| medium | A feature, or a change across several files. | Kimi K3 |
| complex | A whole site, a new journey such as checkout or booking, a rebuild. | GLM 5.2 |
The criterion is the radius of impact, not the length of the message. Any request to build a site or an app from a description is always complex.
If the previous turn failed, the difficulty rises one notch. A question never becomes work. If the classifier does not answer, a keyword rule takes over with no model call at all.
The three models
| Model | Context | What the card says |
|---|---|---|
| Kimi K3 | 1M | Frontier open source model, code leader, native vision. The only one of the three that reads images. |
| GLM 5.2 | 1M | The best open source code model, 1M context, long horizon agentic work. |
| MiniMax M3 | 512K | The generation engine behind Auto. Massive writing at low cost. |
The multiplier next to a model is decorative
The figures shown beside each model are indicative only. Billing goes through the credit calculation, which reads the real consumption reported by the provider. Do not budget on those numbers.
An image overrides your choice, and keeps it
Any message carrying an image goes to Kimi K3. The server enforces it, and the interface records it as your new selection.
Every later message keeps going to Kimi K3 until you pick another model yourself. A notification announces the switch and names the model that took over.
Auto costs latency, not credits
Auto triggers two classifier calls per message. Choosing a model yourself triggers one.
Those calls run before any credit is reserved, and their tokens are never sent to the credit confirmation: only the tokens of the answer itself are. What they cost you is the wait between your click and the first word on screen.
When a model returns nothing
If the engine sends back an empty answer, the reservation is cancelled and you pay nothing. MaShop says so and suggests rewording, or choosing another model in the input bar.
That guard exists because a provider once stopped serving a model still offered in the list. It now covers any cause.