Skip to main content
Krun One is an early model in a closed beta. This page lists what it does not do well yet, so you can design around it.

Tool abstention is advisory

For tool routing (task_type: "tool"), abstain is a hint, not a validated guarantee. On real requests where none of the offered tools applies, Krun One v0 still picks a tool about 39% of the time at the current threshold (BFCL live irrelevance, see Benchmarks). What to do: keep your own fallback for tool calls, such as argument validation, confirmation before side effects, or a “none of these” path in your agent.

Abstention is calibrated only for label-only intents

abstention_status is calibrated only for intent questions whose options have no descriptions. Intent questions with descriptions, and all tool questions, return advisory. See Abstention.

Confidence is a margin, not a probability of being correct

confidence is the top-1 probability minus the top-2 probability. A confidence of 0.4 does not mean the answer is correct 40% of the time. See Confidence and probabilities.

Probabilities are calibrated on benchmark data

Calibration was measured on public intent and function-calling benchmarks. Text that differs a lot from that data, such as other languages, very long documents or highly specialized domains, can have less reliable probabilities.

Four decision primitives, no free text

Krun answers choice, noul, score and multi questions. It does not generate text, images or audio: every answer is a structured decision.

Noul and score calibration depends on the domain

noul probabilities and score distributions were calibrated on the propositions and scales Krun One was evaluated on (intents, tool fit, moderation and response-quality properties, harm severity, sentiment). On very different propositions or texts the ranking usually holds, but probabilities can be over-confident. score was validated mostly on 5-level scales. Check thresholds on your own data.

Cold starts affect latency

The model runs on serverless GPUs that scale to zero. After a period without traffic, the first request can take up to about 2 minutes while the model loads. The API waits up to 150 seconds for the model and otherwise returns 504 UPSTREAM_TIMEOUT. Set client timeouts of at least 180 seconds, retry UPSTREAM_TIMEOUT, and see Production best practices.

Limits per request

Requests over these limits fail with 400 or 413 before any inference runs. See Errors.

Multimodal inputs

Krun One V1 multimodal inputs (images, documents, audio) have their own known limitations:
  • Audio clips longer than 20 seconds may be transcribed only partially.
  • Abstention on media inputs is advisory, not a validated guarantee.
  • Several media items per request work, but their answer quality has not been validated.
  • Numeric comparison and hard visual negatives are weak: validate such questions on your own data.
See Multimodal inputs.

Server-side only

Krun is a server-to-server API. There is no browser SDK, and the API does not send CORS headers.