South KoreaKoreanInstruction tunedOpen weights

Explore KULLM3

Use KULLM3 for Korean conversation, instruction following, explanations, drafting, and knowledge-oriented prompts.

Model

nlpai-lab/KULLM3

Input

Configure your request

text_inputs

Passed directly as TextGenerationPipeline text_inputs. Enter raw text or a JSON chat-message array when the tokenizer has a chat template.

Parameters

return_full_text return_text return_tensors clean_up_tokenization_spaces continue_final_message

handle_long_generation

prefix

Sample outputDisable for greedy, deterministic decoding.

Temperature

Top P

Top K

Min P

Relative probability floor; zero disables it.

Typical P

Maximum output tokens

Input tokens and generated tokens together must fit the 4,096 token context window.

Minimum output tokens

Repetition penalty

No-repeat n-gram size

Beam count

Use KV cache

Length penalty

Early stopping

num_return_sequences