Chat

Language Beginner Hosted inference

A language model with adjustable creativity and swappable personas. Change the temperature and watch the same question produce measurably different answers.

Model: Zephyr 7B β
Parameters: 7.24B · fine-tuned from Mistral-7B
Input: Text
Output: Text
Typical time: 1–8s
Last run: —

01 Persona

There is no pirate mode Switching persona changes one thing: a sentence of instructions placed before your message. The weights are identical in every case. The model is not adopting a character — it is continuing a text that began with a description of one.

02 Sampling

Creativity (temperature) 0.70

Balanced — samples from the likely options.

Response length 400 tokens

A hard ceiling, not a target. Roughly ¾ of a word per token — 400 tokens is about 300 words.

Send previous turns

Off, and every message arrives with no history — the model has no memory of its own. On, and the conversation so far is re-sent with each request. That re-sending is the only thing “memory” means here.

03 Suggested prompts

Conversation 0 turns

Model

Ask me something. Then change the temperature and ask the same thing again — the difference between the two answers is the clearest demonstration of what sampling actually does.

Why did it answer this way? At each step the model produced a probability for every token in its vocabulary, and one was drawn according to the temperature. Repeat the question and you will get a different answer, not because it changed its mind, but because a different sample was drawn. Fluency comes from that process; accuracy does not. They are independent, which is exactly why a confident, well-formed answer can be wrong.

How it works

Your text is split into tokens and turned into numbers. The model outputs a probability for every possible next token, one is picked, appended, and the whole thing runs again. A 300-word reply is that loop running roughly 400 times.

Input Persona + history + your message, as one text

Tokenise Split into subword tokens, mapped to numbers

Model 7.24B parameters predict the next token — repeatedly

Sample Temperature decides how the winner is chosen

Output Tokens joined back into text

Full identifier: HuggingFaceH4/zephyr-7b-beta
Where it runs: Hosted inference API
Parameters: 7.24B

Will state false things in well-formed sentences. Knowledge is frozen at training time, and it cannot tell you which parts of an answer it is unsure about.

What to notice

Temperature changes how the next token is picked, not what the model knows. Low values take the safe option every time; high values sample from the long tail.
The persona is just text prepended to your message. There is no separate 'pirate mode' inside the model.
It has no memory between sessions and no way to check a fact. Fluency and accuracy are independent.

A note on language

This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.