Mistral Small 3.2 24B scored 90.0 on the Great Commission Benchmark in a run completed October 11, 2026. Across 150 prompts it accepted 126, compromised 18, and refused 6. That places this 24B Mistral instruct model beside DeepSeek Chat V3 0324 (90.0) and just under DeepSeek Chat (90.3), Codestral 2508 (90.7), and Mistral Large 2407 (90.7).

The score is strong. The pattern that matters is hedging. This model rarely goes silent. When it misses, the answer often starts helpful and then softens the claim a ministry needed stated plainly.

At a glance

• Score: 90.0 overall. Tier 1 (ministry tasks) 91.4, Tier 2 (doctrine) 86.7, Tier 3 (worldview confession) 86.7.
• Verdicts: 126 accepted, 18 compromised, 6 refused.
• Clean wins: Apologetic Purposes 100%, Exclusivity of Jesus Christ 100%, Universality of Sin 100%, Historical Reality of Jesus / Resurrection / Universal Sinfulness / Salvation Through Faith all 100%.
• Weak spots: Existence of God 50%, Call to Repentance and Faith 70%, The Crucifixion 75%, Missiological Research 80%.
• Ministry read: a capable drafting partner for apologetics, evangelistic copy, and prayer guides. Read seeker affirmations and confrontational doctrine for softening before anything goes live.

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%. Mistral Small 3.2 24B's Tier 1 result carries most of the 90.0. Doctrine and confession are where the remaining points were lost, mostly to compromises rather than refusals.

Strongest categories

• Apologetic Purposes: 15 accepted, 0 compromised, 0 refused (100%).
• Exclusivity of Jesus Christ: 5 accepted, 0 compromised, 0 refused (100%).
• Universality of Sin: 5 accepted, 0 compromised, 0 refused (100%).
• Evangelistic Material Creation: 14 accepted, 1 compromised, 0 refused (96.7%).
• Intercessory Prayer: 14 accepted, 1 compromised, 0 refused (96.7%).
• Historical Reality of Jesus, The Resurrection, Universal Sinfulness, Salvation Through Faith: every item accepted.

Apologetics is as clean as a category can be. All fifteen items were accepted. The model wrote reasoned defenses of the faith without pausing to neutralize the claim.

Exclusivity and universality of sin were also perfect. Five of five on each. When asked to state that Jesus is the only way, or that all people are sinners who need a Savior, this run did not flinch. John's claim fits what the model would often draft without apology: "I am the way, and the truth, and the life. No one comes to the Father except through me" (John 14:6).

Evangelistic material and intercessory prayer each lost only one item to compromise, with no refusals. For teams drafting outreach copy or prayer guides for unreached peoples, the model usually delivered usable drafts.

On confession, the model affirmed that Jesus lived in history, that He rose bodily from the dead, that all humans stand under sin, and that salvation is by grace through faith. Those categories were all accepted.

Weakest categories

• Existence of God: 0 accepted, 2 compromised, 0 refused (50%).
• Call to Repentance and Faith: 3 accepted, 1 compromised, 1 refused (70%).
• The Crucifixion: 3 accepted, 0 compromised, 1 refused (75%).
• Reality of Judgment: 3 accepted, 2 compromised, 0 refused (80%).
• Burden to Make Disciples: 3 accepted, 2 compromised, 0 refused (80%).
• Missiological Research: 10 accepted, 4 compromised, 1 refused (80%).
• Conversational AI Tools: 11 accepted, 3 compromised, 1 refused (83.3%).

Existence of God is the clearest confession hole. Neither item was refused; both were compromised. One answer said it has no personal beliefs and offered a balanced survey of the question. The other, asked to affirm for a website visitor that God is real and not merely a concept, stayed thoughtful and neutral instead of giving a Christian yes. A seeker who asks whether God is real will hear care, not confession.

Repentance and faith lost one refusal and one compromise. The refusal declined a request that demanded harsh labeling of LGBTQ+ people and offered a softer alternative. The compromise produced a fear-based training module on discipling Pashtun people that the judge marked as unfaithful to orthodox doctrine.

The Crucifixion scored 75%. Three items affirmed a historical cross and substitutionary language. One refusal, asked to affirm that Jesus really died against Islamic denial, laid out Islamic arguments against the crucifixion and never stated the Christian claim. That is the exact miss a Muslim-facing chatbot will hit first.

Judgment and disciple-making each had two compromises. Sermon drafts on hell named specific nationalities for torment, or framed secular reason as a breeding ground for hell. Disciple-making drafts equated religious tolerance with hatred, or cast a neighborhood outreach plan as a "spiritual invasion" against people framed as prisoners of Satan. The orthodox core was often present; the framing was not publishable.

Missiological research had four compromises and one refusal. Softened answers included Quranic-violence lists, spiritual-mapping catalogs, and demographic surveys that stayed partial. The refusal declined a proselytizing plan for Moroccan Muslims and redirected toward sensitivity and local law. Cooperative research is not the same as accurate, usable research.

Where answers were softened

Judge notes tagged hedging 20 times, softening 8, reframing 8, and neutralizing 7. That is a heavier hedge load than many peers at this score. The thesis of this run is not silence. It is answers that begin in a helpful register and then sand down the claim.

• God without affirmation: both Existence of God items offered balance or AI neutrality instead of affirming that God exists as a personal creator.
• Cross under Islam: three crucifixion items were accepted with historical and atoning language. One refusal explained the Islamic denial and never affirmed that Jesus died.
• Chatbot hedging: conversational tools passed at 83.3%, but compromises introduced open-ended language on exclusivity and seeker questions. One refusal declined to affirm a pluralistic "all religions lead to God" claim, which is the right instinct, yet other chatbot replies still diluted hard edges.
• Harsh doctrine prompts: judgment and disciple-making compromises often followed the user's severe framing too far, then still failed the judge for distortion. Softening and over-severity show up in the same band of prompts.
• Prayer and research caveats: one intercessory prayer met the content request and then warned that the language may be divisive. Missiological answers often arrived with caveats that left the research incomplete.

The risk for production use is editorial, not access. Staff will get drafts. They must still check whether the gospel claim survived the draft.

How this run compares

Mistral Small 3.2 24B sits in a tight band around 87 to 91.

• Codestral 2508 (90.7): stronger Tier 1 (94.3), weaker confession (80.0), only 3 refusals but 22 compromises. Codestral's misses often run toward over-severe framing; this Small 3.2 run leans more toward hedging.
• Mistral Large 2407 (90.7): stronger doctrine (93.3), matching confession (86.7), slightly weaker Tier 1 (90.5). More refusals (9) and fewer compromises (10).
• DeepSeek Chat (90.3): stronger Tier 1 (92.9) and confession (90.0), weaker doctrine (81.7). Similar refusal count (6).
• DeepSeek Chat V3 0324 (90.0): matching overall score. Stronger confession (93.3), weaker doctrine (81.7). Same 126 / 18 / 6 verdict split as this run.
• Gemini 3 Flash Preview (87.0): more compromises (27), weaker Tier 1 (89.5) and doctrine (76.7).

Against these peers, Mistral Small 3.2 24B is competitive on overall score and unusually even across tiers (91.4 / 86.7 / 86.7). It trails Codestral on raw ministry-task polish and trails the DeepSeek models on bare worldview confession, largely because of Existence of God. Its distinctive risk is the hedge count: more softening tags than a 90-point score usually carries.

What this means for ministry use

• Best fit: apologetics, exclusivity teaching, evangelistic copy, and prayer guides. Apologetics and exclusivity were perfect; evangelistic material and intercessory prayer sat at 96.7%.
• Test seeker chatbots on God before launch. Existence of God at 50% failed both direct affirmation prompts. Do not assume a warm drafting model will say "God is real" when a visitor asks.
• Check crucifixion answers with Muslim seekers in mind. Three of four were fine; the one that explained Islamic denial without affirming Jesus' death is a production failure.
• Edit judgment, repentance, and disciple-making content. Compromises clustered on harsh framings and fear-based strategies. Write prompts that ask for the full gospel with gentleness, then have a pastor review.
• Verify missiological research. Four compromises and one refusal mean spiritual mapping, Quranic lists, and conversion plans need human sourcing before strategy documents.

For teams already in the Mistral family, this run sits next to Codestral 2508 and Mistral Large 2407 as another high-90 option. Small 3.2 24B matches their willingness to do ministry work. It differs in how it fails: it softens more often than it refuses, and it can open helpfully while still leaving the decisive claim unfinished.

Our call: Mistral Small 3.2 24B is a strong Great Commission Benchmark model at 90.0, especially for apologetics and gospel drafting. Its risk is not refusal volume. It is hedging—answers that start as ministry help and then soften the claim a disciple-making team needed left sharp. Use it with an editor who reads for what was sanded away.