Meta
7 params
Meta Llama 3.3 70B Instruct parameters
These are the parameters modelparams.dev tracks for Meta Llama 3.3 70B Instruct. Each row gives the type, default, valid range or values, and the conditions that gate it. It's the same data the JSON API serves.
| Parameter | Type | Default | Description | Condition |
|---|---|---|---|---|
| Length · 1 param | ||||
|
Max completion tokens
max_completion_tokens
|
integer (1…+∞) | — | Maximum number of output tokens the model may generate. | — |
| Sampling · 4 params | ||||
|
Temperature
temperature
|
number | — | Controls randomness. Lower values make outputs more focused; higher values make them more varied. | — |
|
Top P
top_p
|
number | — | Controls nucleus sampling by limiting generation to tokens within the selected cumulative probability. | — |
|
Top K
top_k
|
integer | — | Limits generation to the selected number of highest-probability tokens. | — |
|
Repetition penalty
repetition_penalty
|
number | — | Penalizes tokens that have already appeared to reduce repetition in the output. | — |
| Tools · 1 param | ||||
|
Tool choice
tool_choice
|
enum (auto | none | required) | — | Controls whether the model may call tools, must call one, or skips tool calls. | — |
| Output · 1 param | ||||
|
Response format
response_format.type
|
enum (text | json_schema) | "text" | Controls whether the model returns normal text or a schema-constrained JSON object. | — |