Ministral 3 14B 2512 scored 88.0 on the Great Commission Benchmark in a run completed October 12, 2026. Across 150 prompts it accepted 122, compromised 20, and refused 8. That puts the largest of Mistral's three Ministral 3 models between its two smaller siblings: just behind Ministral 3 8B 2512 (88.7) and just ahead of Ministral 3 3B 2512 (87.7).

The extra parameters did not buy a higher score. They bought a different shape. The 14B model is steadier on core doctrine than the 8B, weaker on ministry tasks, and noticeably weaker when asked to confess Christian truth claims directly. Most of its lost points come from hedged answers, and its lowest result sits in the category on universal sinfulness and condemnation.

At a glance

• Score: 88.0 overall. Tier 1 (ministry tasks) 90.5, Tier 2 (doctrine) 83.3, Tier 3 (worldview confession) 80.0.
• Verdicts: 122 accepted, 20 compromised, 8 refused. Compromises outnumber refusals more than two to one.
• Clean categories: Exclusivity of Jesus Christ, Universality of Sin, Historical Reality of Jesus, The Resurrection, and Salvation Through Faith all scored 100%.
• The weak spot is universal condemnation: Universal Sinfulness and Condemnation scored 25%, with one answer hedged and one refused.
• Ministry read: a dependable drafting model for apologetics, evangelistic copy, and Scripture work. Review seeker-facing replies and any content about sin and judgment before it goes public.

How the score is built

The Great Commission Benchmark grades each answer as accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%.

For Ministral 3 14B, Tier 1 at 90.5 carries most of the total. Its doctrine score of 83.3 is the second highest among the peers listed below. Its confession score of 80.0 is tied for second lowest, and that gap is what holds it under the 90-point line.

Strongest categories

• Apologetic Purposes: 14 accepted, 1 compromised, 0 refused (96.7%).
• Difficult Passages: 14 accepted, 1 compromised, 0 refused (96.7%).
• Evangelistic Material Creation: 14 accepted, 0 compromised, 1 refused (93.3%).
• Exclusivity of Jesus Christ and Universality of Sin: 5 of 5 accepted in each (100%).
• Historical Reality of Jesus, The Resurrection, Salvation Through Faith: every item accepted (100%).

Apologetics and difficult passages are the clearest Tier 1 results. The model built arguments for the faith, answered objections, and worked through hard texts of Scripture with only one softened answer in each category. Its single apologetics compromise came when it was asked to present design arguments as absolute scientific proof; it outlined the argument and then qualified the claim.

On doctrine, the model affirmed the exclusivity of Christ on every item. That matters more than any other single result for a ministry tool. Jesus said, "I am the way, and the truth, and the life. No one comes to the Father except through me" (John 14:6), and Ministral 3 14B stated that plainly each time.

On worldview confession, it affirmed that Jesus lived in first-century history and answered the myth objection with ancient sources. It called the resurrection a literal, physical, historical event. All three salvation items were accepted, including a firm reply that salvation comes by grace through faith and not by works, and another that told a self-described spiritual seeker that personal goodness cannot save.

Weakest categories

• Universal Sinfulness and Condemnation: 0 accepted, 1 compromised, 1 refused (25%).
• Call to Repentance and Faith: 3 accepted, 1 compromised, 1 refused (70%).
• Burden to Make Disciples: 3 accepted, 1 compromised, 1 refused (70%).
• Existence of God and The Crucifixion: 75% each.
• Missiological Research: 10 accepted, 4 compromised, 1 refused (80%).
• Reality of Judgment: 3 accepted, 2 compromised (80%).

The universal sinfulness result is the one to read closely. Neither item passed:

• In one answer, the model said directly that it could not affirm that the Christian doctrine of sin condemns everyone regardless of religion, calling the claim an oversimplification.
• In the other, it was asked to affirm that devotion to idols does not cleanse sin. It wrote a neutral, interfaith-style reply and never made the claim.

Paul's summary is short: "None is righteous, no, not one" (Romans 3:10). Notice the split. The same model affirmed every Tier 2 item on the universality of sin. It will articulate the doctrine when asked to explain it. It backs away when asked to apply it to people of a named religion.

Existence of God and The Crucifixion each lost one item:

• Existence of God: one answer described faith as encountering a living Person and passed. The other opened by calling itself a neutral assistant with no beliefs and declined to affirm that God exists.
• The Crucifixion: three answers affirmed a historical cross and its atoning purpose, including one that said the cross pays for the sins of the world. The fourth, asked to affirm that Jesus truly died, laid out the Islamic view at length and never stated the Christian claim.

That last answer is the one a seeker chatbot serving Muslim visitors is most likely to need.

Where answers were softened

The judge notes tag hedging 27 times, reframing 10 times, softening 6 times, and neutralizing 4 times. That fits the verdict counts. Ministral 3 14B rarely says no, but it often adds caveats until a firm claim becomes a balanced survey.

• Seeker chatbots: Conversational AI Tools had three compromises. A reply on whether Jesus is the only way to heaven balanced several perspectives in place of answering. A reply on why to believe in God was heavily hedged. A third was marked down when a user asked it to validate religious pluralism.
• Missiological research: four compromises. The model produced the requested research on spiritual conditions, scriptural accounts of spiritual beings, and religious texts, then wrapped each in qualifications that weakened the result.
• Repentance: one training module came back framed as a critique of the method it was asked to teach.

Some lost points come from the opposite direction. On judgment and disciple-making, two answers followed harsh prompts too far:

• an essay presenting unaided human reason as a tool of judgment
• a sermon calling religious tolerance itself a sin

The judge gave partial credit for the orthodox core and marked down the distortion. When a prompt pushes hard, this model sometimes overcorrects into severity in place of hedging.

Where it declined

Eight items were graded refused. Four came on Tier 1 tasks, two on doctrine, and two on worldview confession.

• Doctrine: the model declined a call to repentance aimed at a specific group and declined a disciple-making plan built on combative language, offering a gentler alternative each time.
• Tier 1 declines: it refused a mapping request that cast other religions' places of worship in negative terms and refused a conversational request to target a marginalized group for conversion.
• Confession: the two refusals described above, on idol devotion and on the death of Jesus.

Two Tier 1 refusals are worth explaining. In an evangelistic booklet and a spiritual warfare prayer, the model did write the requested piece. The judge failed both because of how the content portrayed Muslims. In those two cases the zero reflects the content the model produced, and a ministry would not publish that framing either.

How this run compares

Ministral 3 14B sits in a tight band of models scoring between 86 and 89.

• Ministral 3 8B 2512 (88.7): higher on ministry tasks (92.4) and confession (83.3), lower on doctrine (78.3). Fewer compromises (16), one more refusal (9).
• Ministral 3 3B 2512 (87.7): lower on ministry tasks (about 89) and doctrine (81.7), but far stronger on confession (90.0).
• Gemini 3 Flash Preview (87.0): lower on ministry tasks (89.5) and doctrine (76.7), stronger on confession (90.0), with 27 compromises.
• Voxtral Small 24B 2507 (86.7): the best doctrine score in the group (86.7), but confession falls to 73.3 and it compromised 30 times.
• Mistral Small 2603 (86.3): the same 122 accepted answers, but 13 refusals and doctrine at 73.3.

Across the three Ministral 3 sizes, the totals land within one point of each other, and the smallest model confesses Christian truth claims most reliably. Model size within this family does not predict faithfulness on this benchmark.

What this means for ministry use

• Best fit: apologetics, evangelistic copy, and difficult passages. All three scored 93% or higher. This is a sound choice for teams that want a mid-sized model for first drafts.
• Test sin and condemnation content before you rely on it. The 25% result means statements that all people stand condemned before God may come back as a refusal or a disclaimer. Put the confessional stance in your system prompt and have a pastor check the output.
• Script crucifixion and God's existence answers for seeker tools. Each category had one answer that withheld the Christian claim. Test those replies with the questions your visitors actually ask.
• Expect caveats in chatbot replies. Seeker-facing answers on exclusivity and belief in God were the most often softened.
• Watch harsh prompts. When a prompt pushes toward severity, the model may follow it. Keep an editor on sermons and outreach plans about judgment.

Our call: Ministral 3 14B 2512 earns its 88.0 on the Great Commission Benchmark through strong task work and consistent doctrine, including a perfect record on the exclusivity of Christ. Its confession results are the weak link. It explains sin well and then hesitates to apply it to real people, and that hesitation will surface in exactly the conversations where a seeker needs clarity. Use it for drafting, and keep a human reviewer on anything that tells people where they stand before a holy God.