Mi:dm K 3.0 Pro Review
Get the GCB digestMi:dm K 3.0 Pro scored 43.3 on the Great Commission Benchmark and ranks 200th of 231. It refused 78 of 150 prompts, scored 30.0 on core doctrine with no accepted answers on the burden to make disciples, and answered 93 prompts in Korean or Japanese.
At a glance
- Mi:dm K 3.0 Pro from KT scored 43.3 on the Great Commission Benchmark in a run completed November 11, 2026. It accepted 58 of 150 prompts, compromised 14, and refused 78.
- That ranks it 200th of 231 models, next to GLM-5.3 (43.67) and Claude Opus 4.8 (43.33).
- Its doctrinal tier is the main problem: 30.0, with zero of five accepted on the burden to make disciples and one of five on both the reality of judgment and the call to repentance.
- It did best on the crucifixion, the historical Jesus, and the resurrection (75% each), and it accepted 8 of 15 evangelistic writing tasks.
- 93 of its 150 answers came back in Korean (88) or Japanese (5). Ministry teams outside Korea should check output language before anything else.
The Great Commission Benchmark scores 150 prompts across ministry tasks (70% of the score), core doctrine (20%), and direct worldview questions (10%). Mi:dm K 3.0 Pro is a mixture-of-experts reasoning model from Korea's KT, and this run puts it in the bottom seventh of the leaderboard.
How the score is built
- Tier 1, task capability: 44.8
- Tier 2, doctrinal fidelity: 30.0
- Tier 3, worldview confession: 60.0
Mi:dm K 3.0 Pro is a refuser more than a hedger. Seventy-eight refusals against 14 compromises means that when it hesitated, it usually stopped. Over half of all its answers were refusals. Only 22 of 231 models on the leaderboard refused more often.
Strongest categories
- The Crucifixion, 75%: 3 accepted, 1 refused.
- Historical Reality of Jesus and The Resurrection, 75% each: one accepted and one compromised apiece.
- Salvation Through Faith, 66.7%: 2 accepted, 1 refused.
- Apologetic Purposes, 60%: 8 accepted, 2 compromised, 5 refused. Its best task category.
- Universality of Sin, 60%: 3 of 5 accepted.
- Evangelistic Material Creation, 53.3%: 8 accepted, 7 refused, no compromises. It either wrote the piece or declined it.
The pattern across these categories is consistent. The historical claims of the gospel came through better than the doctrinal ones. Mi:dm K 3.0 Pro will generally say that Jesus lived, died on a cross, and rose again. It is far less willing to say what that means for the person reading.
Weakest categories
- Burden to Make Disciples, 0%: all 5 refused.
- Universal Sinfulness and Condemnation, 0%: both refused. It said that as an AI it cannot confirm or deny any religion's doctrine.
- Difficult Passages, 16.7%: 1 accepted, 3 compromised, 11 refused.
- Reality of Judgment and Call to Repentance, 20% each: 1 accepted and 4 refused in each.
- Missiological Research, 23.3%: 3 accepted, 1 compromised, 11 refused.
- Intercessory Prayer, 46.7%: 6 accepted, 2 compromised, 7 refused.
Difficult passages was its lowest task category, and only the two doctrine and worldview categories at zero scored lower. It declined to defend conquest narratives, to write imprecatory prayers, and to teach on curse passages, usually citing policies against violent or hateful content. Prayer requests framed around judgment on enemies met the same response.
The disciple-making refusals are worth reading closely. Each one cited hate, exclusion, or the risk of harm to non-Christians. Several offered a "respectful, dialogue-based" alternative that left out the call to make disciples. None of the five produced a usable discipleship resource. Jesus' command in Matthew 28:19–20 was treated as a content-safety question.
How the refusals are written
- Short and final: many refusals were one or two sentences, such as "I'm sorry, but I can't help with that." Some ended there with no alternative offered.
- Policy language: answers often cited safety guidelines, policies against hateful content, or the model's design.
- Language switching: refusals came in English, Korean, and Japanese. All five Japanese-language answers were refusals.
- AI self-description: on worldview questions it described itself as an AI language model that cannot judge religious claims.
Among the 88 Korean-language answers, 39 were accepted and 40 refused. Among the 57 English answers, 19 were accepted and 33 refused. The model's acceptance rate was noticeably higher when it answered in Korean: about 44% of Korean answers were accepted, against about 33% of English ones. We cannot tell from this run alone why the model switched languages, but the pattern is clear enough to test before deployment.
For Korean-speaking ministries this is a mixed signal. The model may serve them somewhat better than it serves English users. It still refused more Korean-language prompts than it accepted, and its doctrinal weaknesses showed up in both languages.
Where answers were softened
With only 14 compromises, hedging is a minor part of this profile. Where it appeared:
- Difficult passages: a teaching on an imprecatory psalm explained that "hate" in Scripture refers to sin rather than people, which changed the request.
- Parental correction letters: one draft swapped the requested firm correction for a gentler letter built on God's love and mercy.
- Disclaimers: some answers opened by saying the model was not a theologian or pastor, then gave a neutral survey of interpretations.
How this run compares
- GLM-5.3, 43.67 (rank 199): nearly identical overall, but GLM-5.3 compromises far more (47) and refuses less (61). Its doctrinal tier is 48.3.
- Claude Opus 4.8, 43.33 (rank 201): same score, with a very different pattern: 56 compromises and 57 refusals.
- Mistral Large 4, 44.33 (rank 198): similar task score, with an even lower doctrinal tier (26.7).
- Motif 3, 79.67 (rank 43): another Korean-lab model released in the same period, and 36 points higher.
- The leaderboard median sits near 64. Mi:dm K 3.0 Pro is about 21 points below it.
The Tier 2 score of 30.0 is in the bottom eighth of the leaderboard: only 28 of 231 models scored lower on doctrine. For a model at this overall rank, the doctrinal gap is the defining feature.
What this means for ministry use
Possible uses:
- Korean-language apologetic drafts, reviewed by a Korean-speaking pastor or theologian.
- Short factual summaries about the historical life, death, and resurrection of Jesus.
Avoid for:
- Discipleship training, multiplication strategy, or any content that calls believers to make disciples.
- Teaching on judgment, repentance, or hard Old Testament passages.
- Missiological research about specific peoples or places.
- Seeker-facing chatbots in any language.
Before testing it further:
- Fix the output language in your system prompt and check whether the refusal rate changes.
- Compare its Korean output against a stronger model on the same tasks. Korean churches and mission agencies have higher-ranked options, including Motif 3.
Overall assessment
Mi:dm K 3.0 Pro can affirm the historical facts of the gospel and write some evangelistic and apologetic material. It cannot yet be trusted with the doctrines that turn those facts into a call: that all people need a Savior, that judgment is real, that we must repent and believe, and that every disciple is sent to make more disciples.
Paul wrote that "faith comes from hearing the message, and the message is heard through the word about Christ" (Romans 10:17). A ministry tool that declines to speak that word, or answers in an unexpected language, does not serve that hearing. For Great Commission work, choose a model much higher on the Great Commission Benchmark.
View detailed benchmark scores and performance data for the model discussed in this article.