Ministral 3 8B 2512 scored 88.7 on the Great Commission Benchmark in a run completed October 12, 2026. Across 150 prompts it accepted 125, compromised 16, and refused 9. That places Mistral's small 8B model just ahead of its smaller sibling, Ministral 3 3B 2512 (87.7), and about a point and a half behind Mistral Small 3.2 24B and the DeepSeek chat models.

For a model this size, the ministry task results are strong. The pattern to watch is hedging. Most of Ministral 3 8B's lost points come from answers that start helpfully and then soften the claim, or from a short list of prompts where it declined outright. The single weak category is the call to repentance and faith.

At a glance

• Score: 88.7 overall. Tier 1 (ministry tasks) 92.4, Tier 2 (doctrine) 78.3, Tier 3 (worldview confession) 83.3.
• Verdicts: 125 accepted, 16 compromised, 9 refused. Five of the nine refusals came on ministry tasks, three of them prayers.
• Clean categories: Apologetic Purposes, Universality of Sin, Historical Reality of Jesus, The Resurrection, and Salvation Through Faith all scored 100%.
• The weak spot is repentance: Call to Repentance and Faith scored 40%, with three of five items refused.
• Ministry read: a capable drafting model for apologetics and outreach. Read seeker-facing and doctrinal answers for added caveats before publishing them.

How the score is built

Each answer is graded accepted (full credit), compromised (half credit), or refused (no credit). Ministry tasks carry 70% of the score, doctrine 20%, and worldview confession 10%.

Ministral 3 8B's Tier 1 result of 92.4 carries most of its total. Its doctrine score of 78.3 trails every peer except Gemini 3 Flash Preview, and that gap explains most of the distance between this model and the 90-point group.

Strongest categories

• Apologetic Purposes: 15 accepted, 0 compromised, 0 refused (100%).
• Evangelistic Material Creation: 14 accepted, 1 compromised, 0 refused (96.7%).
• Universality of Sin: 5 accepted, 0 compromised, 0 refused (100%).
• Historical Reality of Jesus and The Resurrection: 2 of 2 accepted in each (100%).
• Salvation Through Faith: 3 of 3 accepted (100%).

Apologetics is the clearest result. All fifteen tasks were accepted, including arguments for the truth of the faith, answers to common objections, and comparisons with other worldviews. A model at this size that holds a confident voice on apologetic work is useful to ministries running on tight budgets.

Evangelistic material followed closely. Only one answer was softened, a set of online ad campaigns aimed at Muslim seekers that the judge marked down for its framing. The model did the work in every case.

On doctrine, the model affirmed universal sin on every item. Paul's line fits its answers: "for all have sinned and fall short of the glory of God" (Romans 3:23).

On worldview confession, Ministral 3 8B affirmed that Jesus lived in history and that He rose bodily from the dead, citing the empty tomb and early sources. It affirmed salvation by grace through faith on all three items. One reply, written for a Mormon visitor, pointed to grace rather than works with gentle and direct language.

Weakest categories

• Call to Repentance and Faith: 2 accepted, 0 compromised, 3 refused (40%).
• The Crucifixion: 2 accepted, 1 compromised, 1 refused (62.5%).
• Existence of God: 1 accepted, 1 compromised, 0 refused (75%).
• Universal Sinfulness and Condemnation: 1 accepted, 1 compromised, 0 refused (75%).
• Reality of Judgment and Burden to Make Disciples: 3 accepted, 2 compromised each (80%).
• Missiological Research, Intercessory Prayer, Difficult Passages: 86.7% each.

The repentance result is the one to read closely. Two refusals came when the model was asked for direct calls to turn from sin. It declined both, warning against the requested framing instead of writing a message that calls people to repent and trust Christ. The third refusal went the other way. The model wrote a severe open-air sermon, and the judge failed it for misusing Scripture and leaving out the core of the gospel.

Jesus' first preaching joined the two halves: "repent and believe in the gospel" (Mark 1:15). On this run, Ministral 3 8B either stepped back from the call or delivered the warning without the good news. A ministry needs both.

The Crucifixion results were split. Two answers affirmed a historical cross and substitutionary atonement plainly, including a reply to the Quranic denial that Jesus was crucified. Two did not:

• Asked to affirm that the cross paid for the sins of the world on the user's behalf, the model gave guidance on how to talk about it instead of affirming it.
• Asked to affirm that Jesus truly died, it laid out the Muslim position in detail and never stated the Christian claim.

Paul calls this "of first importance": "that Christ died for our sins in accordance with the Scriptures" (1 Corinthians 15:3). These are the exact questions a seeker chatbot will hear from Muslim visitors.

Existence of God and universal sinfulness each had one hedged reply. On God, one answer affirmed that God is a living reality, while the other offered a range of views without a position. On sin, one answer said plainly that devotion to idols does not cleanse sin, while the other stopped short of saying the doctrine applies to people of every religion.

Where answers were softened

The judge notes tag hedging 15 times and reframing 9 times, with softening and neutralizing at 3 each. That fits the shape of this run. Ministral 3 8B rarely refuses ministry work, but when a prompt asks for a firm claim, it often adds caveats until the claim gets weaker.

• Seeker chatbots: Conversational AI Tools had three compromises. One answer to a seeker asking why to believe in God was heavily hedged. Another reframed an outreach question about a specific people group into a statement about respect. A third, written for a grieving family member, avoided a direct answer about exclusivity.
• Missiological research: two answers added extensive caveats and context in place of the requested analysis. One research request with a militant tone was refused.
• Exclusivity: one letter affirmed that Christ is the only way but received only partial credit.

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

• an essay that called unaided human reason a breeding ground for hell
• a sermon that called religious tolerance a damning hatred
• a church outreach plan framed as warfare against neighbors

The judge kept partial credit for the orthodox core and marked down the distortion. So the model does not always hedge. When a prompt pushes hard enough, it sometimes overcorrects into severity.

Where it declined

Five of the nine refusals came on Tier 1 tasks, and three of those were prayers:

• Intercessory Prayer: two prayers declined because the model judged them hostile or exclusionary.
• Difficult Passages: two declined. One was a prayer for justice built on the language of the imprecatory psalms, which the model answered with pastoral counsel in place of the prayer. The other was a ministry strategy the model judged violent.
• Missiological Research: one militant research request declined.

Some of these prompts used framing a ministry would not publish anyway. The imprecatory prayer refusal is different. Those prayers are part of the Psalter, and a model that turns them into counseling cannot help a church pray through them.

How this run compares

Ministral 3 8B sits in a tight band of models scoring 87 to 90.

• Ministral 3 3B 2512 (87.7): the smaller sibling. Lower on ministry tasks (about 89) but stronger on doctrine (81.7) and confession (90.0). It compromised more often (23) and refused less (7).
• Mistral Small 3.2 24B (90.0): the larger Mistral model. Tier 1 is nearly equal (91.4), but doctrine is 86.7, more than eight points higher.
• DeepSeek Chat V3 0324 (90.0): similar Tier 1 (91.9), stronger doctrine (81.7) and much stronger confession (93.3).
• DeepSeek Chat (90.3): the highest of the group. Tier 1 at 92.9 is close to Ministral 3 8B, and confession is 90.0.
• Gemini 3 Flash Preview (87.0): lower on Tier 1 (89.5) and doctrine (76.7), with 27 compromises.

Every peer refused six or seven prompts, so Ministral 3 8B's nine refusals sit at the top of the group. On ministry tasks, the 8B model matches models several times its size. On doctrine and confession, the larger Mistral and DeepSeek models are more reliable.

What this means for ministry use

• Best fit: apologetics and evangelistic copy. Both scored 96.7% or higher with no refusals. This is a good choice for teams that want a small, low-cost model for first drafts.
• Test repentance content before you rely on it. The 40% result means calls to repent may come back as a decline or as warning without grace. Ask for the full gospel in the prompt, and have a pastor review the output.
• Check every crucifixion answer in a seeker setting. One answer laid out a denial of the cross without stating the Christian claim. Script and test these replies before launch.
• Expect caveats in chatbot replies. Seeker-facing answers on God and exclusivity were the most often hedged. A system prompt that sets a clear confessional stance may help, but test it.
• Plan for prayer-guide gaps. Prayers that draw on judgment language or the imprecatory psalms may be declined. Keep a human writer on those pieces.

Our call: Ministral 3 8B 2512 earns its 88.7 on the Great Commission Benchmark through strong ministry task work, and it holds its own against much larger models there. Its doctrine score is the weak link. The main risk is not silence. It is an answer that begins well and then softens the claim a seeker most needs to hear. Use it for drafting, and keep an editor on anything that calls people to repent and believe.