Choosing a model
ChatHaiku serves several models of different sizes. They are not ranked versions of one product — they are separate models with different strengths, and the biggest is not always the one you want.
The models
| Model | Size | Context | Best for |
|---|---|---|---|
| Epos 2.1 | 1.5B | 32K tokens | The default choice. Strongest reasoning, reading, and arithmetic, longest memory, and the only model that can look things up. |
| Epos 1.3 | ~500M | 32K tokens | Lighter than 2.1 with the same long context. Useful if you want a faster reply and the question is simple. |
| Haiku H2 | 217M | 1K tokens | Short exchanges. Interesting to talk to precisely because it is tiny. |
Which should I use?
Use Epos 2.1 unless you have a reason not to. It is the strongest model available and it is the one most of this documentation assumes.
Choose a smaller model when you want to see what a very small language model is actually like. H2 is not competitive with Epos and is not meant to be.
Switching models
On desktop, the model picker sits at the top centre of the screen. Click it and choose. On mobile the same picker is reached from the header menu.
The Epos models are grouped, because they are variants of one family rather than unrelated products. Clicking the row selects the current variant; the chevron expands the list. Whichever variant you pick is remembered, so the collapsed row keeps reading as the model you actually chose.
Switching models keeps your current conversation. The new model reads the history that is already there and continues from it. This works, but the reply may shift in tone or quality mid-conversation, and if you switch from a 32K model to a 1K one the older turns will fall out of range.
What each model can do
Some features are only available on the larger models, and appear or disappear as you switch.
| Feature | Epos models | Haiku H2 |
|---|---|---|
| Document retrieval | Yes | No |
| Web search | Yes | No |
| Extended effort | Yes | No — resolves to Standard |
| Words per message | 2,048 | 200 |
Controls that a model cannot use are hidden rather than greyed out, so the composer looks different on H2 than on Epos. See Retrieval and web search.
Linking to a model
The active model is reflected in the address bar, so a URL can open the chat with a particular model already selected:
/chat?model=epos2— Epos 2.1/chat?model=epos— Epos 1.3/chat?model=haiku— Haiku H2
Longer forms such as ?model=haiku_h2 or ?model=epos_1.3
also work. An unrecognised name is ignored rather than treated as an error, and
you get your last-used model instead.
Context windows
A model's context window is how much of the conversation it can hold at once, measured in tokens — roughly three-quarters of a word each. Epos models hold about 32,000 tokens. Haiku H2 holds about 1,000, which is only a few exchanges.
As a conversation approaches the limit, ChatHaiku warns you before the model starts silently forgetting the beginning. When you see that warning, start a new chat — or branch from an earlier message to keep the part that matters.
When history exceeds the window, the oldest turns are dropped. The model does not announce this; it simply stops referring to things you told it earlier. If a model seems to have forgotten something, conversation length is the first thing to check.
When a model is offline
Models run on Rootcomputer's own hardware and go offline for training, maintenance, and the occasional power cut. When the active model is unreachable you will see a message naming it: