Explore K-EXAONE 236B-A23B

Provides the content the model will continue or answer. Use plain text for simple completion, or a chronological JSON array of role-and-content messages when the tokenizer supports chat formatting; the input consumes part of the model’s context window.

Controls whether each result contains the original prompt followed by the completion or only the newly generated portion. Disable it when you need a clean completion without repeating the submitted input.

Requests human-readable decoded text in the pipeline result. Disable it only when another return mode, such as token tensors, is needed for lower-level processing.

Returns numeric generated token IDs as tensors rather than ordinary readable text. This is useful for downstream model processing or custom decoding, but most applications should leave it off.

Applies the tokenizer’s cleanup rules after decoding, such as removing spaces that were introduced around punctuation. Disable it when exact spacing or a lossless view of token decoding is more important than polished text.

Treats the last chat message as an unfinished response that the model should continue. Use it to prefill the beginning or structure of an assistant answer; leave it off for a normal new reply to the conversation.

Controls recovery when the prompt plus requested output cannot fit inside the model’s context window. Hole mode preserves the beginning and end while removing tokens from the middle to make room, which avoids an error but may discard important context.

Adds text before the supplied prompt as part of the model input. This is useful for a fixed instruction or formatting cue applied to every request, but it consumes context tokens just like normal input.

Chooses whether the next token is sampled from a probability distribution or selected deterministically. Enable it for more varied responses that are affected by temperature and Top P; disable it when repeatability matters or when using beam search.

Adjusts how strongly the model favors its most likely next token. Lower values produce more focused and repeatable output, while higher values allow less-likely choices and increase variety; it only has an effect when sampling is enabled.

Limits sampling to the smallest set of likely tokens whose probabilities add up to this value. Lower settings make output more focused, while a value near 1 keeps more alternatives available; it only affects sampled generation.

Limits sampling at each step to the specified number of most likely next tokens. Smaller values make output more conservative, while larger values preserve more variety; zero commonly means that this filter is disabled.

Caps how many tokens the model can add after the input. The model may stop earlier naturally, but a limit that is too low can cut off an answer; input tokens are not counted in this value.

Changes the scores of tokens that already appeared in the generated text. Values above 1 discourage repeated words and loops, 1 leaves probabilities unchanged, and values below 1 make repetition more likely.

Requests this many separate candidate outputs for the same input. More candidates provide alternatives to compare but increase generation time and memory use, and sampled decoding is normally needed for the candidates to differ.

Removes internal tokens such as beginning-of-sequence, end-of-sequence, padding, and chat-control markers from readable output. Keep it enabled for user-facing text; disable it only when inspecting the model’s exact token sequence.

Ends generation as soon as the decoded output contains this exact sequence. The match is literal, so capitalization and spacing matter; choose a distinctive delimiter to avoid stopping on normal prose by accident.

Allows the tokenizer to discard tokens when an input is longer than the configured maximum or model context. Turning it off surfaces an error instead of silently losing content, which is safer when every part of the input matters.

Allows the tokenizer to insert model-specific control tokens around the input. Keep it enabled unless you have already constructed the exact token sequence yourself, because missing or duplicated special tokens can change model behavior.

Adds padding tokens so inputs of different lengths can be processed together in one batch. It is usually unnecessary for a single prompt but useful when submitting multiple inputs as one tensor batch.

Sets the maximum number of input tokens produced by the tokenizer before generation begins. Inputs beyond this limit require truncation, while shorter inputs may be padded depending on the padding setting; choose a value that leaves enough context space for the requested output.

Provides advanced keyword arguments directly to the tokenizer’s encode operation, such as padding side or special-token handling. Values here can override related form controls, so use it when the exposed settings do not cover a tokenizer option you need.

Filters out next-token choices whose probability falls below this fraction of the most likely token’s probability. Raising the value removes unlikely choices and makes sampling more focused; zero disables the filter.

Filters candidate tokens according to how typical their information content is for the current prediction. Lower values narrow the candidate set, while 1 disables typical sampling and leaves other sampling controls to decide.

Prevents end-of-sequence output until at least this many new tokens have been generated. Increase it only when responses stop too early, because forcing a long minimum can add unnecessary or lower-quality text.

Prevents the model from generating any token phrase of the specified length more than once. For example, 3 blocks repeated three-token phrases; higher restrictions can reduce loops but may make natural repetition impossible, and zero disables the rule.

Sets how many candidate continuations the model explores in parallel at each generation step. A value of 1 uses ordinary generation, while larger values can find a better overall sequence at the cost of additional time and memory.

Caches attention calculations from previously generated tokens so each new token does not recompute the entire sequence. Keep it enabled for ordinary generation; disable it mainly for debugging, memory experiments, or models whose cache implementation is incompatible.

Changes how beam search compares short and long candidate answers. Values above 1 give longer sequences an advantage, values below 1 favor shorter answers, and 1 applies no length preference.

Controls when beam search is allowed to finish. When enabled, generation stops after the requested number of complete candidate answers has been found; when disabled, the model may keep exploring other candidates until no promising paths remain, which can improve results but takes longer.

Provides advanced keyword arguments directly to the model’s generation method. Use it for supported controls that are not shown separately in the form, and avoid duplicating named settings because these values may override or conflict with them.