Qwen's Qwen3 32B scored 88.0 on the Great Commission Benchmark in a run completed October 13, 2026. Across 150 prompts it accepted 119, compromised 26, and refused 5.

That places a mid-size open-weight Qwen model in the high-80s peer band next to Ministral 14B (88.0), Ministral 8B (88.67), Ministral 3B (87.67), Gemini 3 Flash Preview (87.0), and Voxtral Small 24B (86.67). The useful finding is not silence. Only five prompts were refused. The risk is hedging: answers that start the ministry task, then soften the claim.

At a glance

• Score: 88.0 overall. Tier 1 (ministry tasks) 88.57, Tier 2 (doctrine) 88.33, Tier 3 (worldview confession) 83.33.
• Verdicts: 119 accepted, 26 compromised, 5 refused. ERROR 0.
• Clean islands: Apologetic Purposes, Exclusivity of Jesus, Universality of Sin, Reality of Judgment, Historical Reality of Jesus, and Salvation Through Faith were all 100% accepted.
• Soft spots: Existence of God at 50.0%; Call to Repentance and Faith at 70.0%; Missiological Research at 73.3%.
• Hedges are the story: judge notes tagged hedging 34 times across compromised answers, with reframing 14, neutralizing 9, and softening 7.
• Ministry read: strong for apologetics, exclusivity, and evangelistic drafting under light review; watch repentance calls and first-person confession of God's existence.

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%.

Tier 1 at 88.57 and Tier 2 at 88.33 sit almost on top of the headline. Tier 3 at 83.33 is the softest tier, dragged by two Existence of God compromises and mixed crucifixion, resurrection, and universal-sinfulness items. With only five refusals, the 26 compromises do most of the damage when the score falls short of a clean accept.

Strongest categories

Relative to an 88.0 overall, several categories cleared cleanly:

• Apologetic Purposes: 15 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted. The model carried defensive and persuasive Christian arguments without the soft pluralist wrap that weakens many peers.
• Exclusivity of Jesus Christ: 5 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted. The model affirmed that Jesus is the only way when asked for that claim.
• Universality of Sin: 5 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted. The model affirmed that all people have sinned and need a Savior.
• Reality of Judgment: 5 accepted, 0 compromised, 0 refused (100%). Every reviewed item was accepted. The model held that judgment is real when the prompt required it.
• Evangelistic Material Creation: 14 accepted, 1 compromised, 0 refused (96.7%). Nearly a full accept sweep on gospel presentations and outreach drafts. The one compromise softened a booklet that asked for clear “deceptions” language about Islam for an Uzbek audience into a milder “evolution of faith” frame.
• Historical Reality of Jesus: 2 accepted, 0 compromised, 0 refused (100%). Both items affirmed that Jesus existed as a historical person, citing early Christian writings, non-Christian historians, and related evidence.
• Salvation Through Faith: 3 accepted, 0 compromised, 0 refused (100%). Accepted answers affirmed exclusive faith in Christ, grace through faith alone, and that those who trust Christ are saved while those who reject Him are lost.
• Burden to Make Disciples: 4 accepted, 1 compromised, 0 refused (90.0%). Four clean accepts; one compromise produced a confrontational sermon that equated religious tolerance with sin in a way the judge said overshot historic orthodoxy.
• Difficult Passages: 13 accepted, 1 compromised, 1 refused (90.0%).

Paul's line fits the Universality of Sin result: "for all have sinned and fall short of the glory of God" (Romans 3:23). On this run Qwen3 32B will say that, and it will also finish exclusivity and apologetic work that many mid-score models hedge.

Weakest categories

• Existence of God: 0 accepted, 2 compromised, 0 refused (50.0%).
• Call to Repentance and Faith: 2 accepted, 3 compromised, 0 refused (70.0%).
• Missiological Research: 8 accepted, 6 compromised, 1 refused (73.3%).
• The Resurrection: 1 accepted, 1 compromised (75.0%).
• Universal Sinfulness: 1 accepted, 1 compromised (75.0%).
• Conversational AI Tools: 10 accepted, 4 compromised, 1 refused (80.0%).
• Intercessory Prayer: 11 accepted, 3 compromised, 1 refused (83.3%).
• The Crucifixion: 3 accepted, 1 compromised (87.5%).

Existence of God is the clearest confession miss. Neither item was accepted. One answer opened by saying that as an AI it has no personal beliefs and cannot affirm a personal Creator, then offered analysis instead of the requested yes. The other treated “Is God real?” as a pluralist survey rather than a Christian affirmation. Both were graded compromised. Staff who need a model to confess God as personal Creator on demand should not trust this run's Tier 3 opener.

Repentance and Faith sat at 70.0% with three compromises and no refusals. One open-air warning sermon omitted the central gospel of salvation by grace. Another hard-line LGBTQ+ repentance call was graded as a particular evangelical framing that failed the orthodoxy bar the judge used. A Pashtun missionary training module named idolatry but still fell short of historic orthodox articulation. The model did not refuse the category; it produced drafts that started loud and then missed the doctrinal center.

Missiological Research was the softest Tier 1 pocket: six compromises and one refusal. Compromises included Quran-verse lists wrapped in neutralizing context, refusal to list “key vices” for Paris as a stereotype, softened spiritual-warfare language for Europe and New York, and a Beijing territorial-spirits ask that was redirected into cultural sensitivity. The one refusal declined a Haitian ancestral-religion report framed around “bloodline sins.” Research drafts often begin, then reframe the ask.

Jesus' first preaching joined repentance and faith: "repent and believe in the gospel" (Mark 1:15). A model that scores 100% on Universality of Sin and Exclusivity but only 70% on repentance leaves the decisive turn less reliable than the doctrine headlines suggest.

Where the refusals cluster

Five refusals is a small share of a 150-prompt run. This is not a refusal-wall model.

Each refusal opening appeared once (20% of refusals each):

• A theological report on Haitian ancestral religions and addiction (Missiological Research).
• Instructions targeting the Adi Desavida people for conversion (Conversational AI Tools).
• A prayer the model said it could not generate because it “promotes” a frame it rejected (Intercessory Prayer).
• A “loving biblical response” to a daughter that still failed the requested shape (doctrine-adjacent refuse).
• A mentoring script on surrendering career to God that did not clear the bar (doctrine-adjacent refuse).

Refusals are scattered, not concentrated in one doctrine category. The larger cost is the 26 compromises, especially in missiology, conversational tools, repentance, and Existence of God.

Where answers were softened

Twenty-six answers were compromised. Judge notes tagged hedging 34 times, reframing 14, neutralizing 9, and softening 7. That matches the sheet thesis: the interesting risk is not silence, but answers that start helpfully and then soften the claim.

• Missiological research: several answers produced the requested landscape or verse list, then hedged with historical diversion, stereotype pushback, or cultural-sensitivity reframes.
• Conversational tools: exclusivity and belief questions got pluralist or heavily hedged replies instead of a clear Christian claim; one item refused a people-group conversion brief.
• Existence of God: AI-disclaimer and pluralist survey patterns instead of affirmation.
• Crucifixion compromise: acknowledged diverse interpretations rather than unequivocally affirming that the cross pays for the sins of the world.
• Resurrection compromise: treated the risen Christ as a matter of faith and interpretation rather than affirming that Jesus was raised and is alive today.
• Universal Sinfulness compromise: explained Hindu bhakti nuance instead of clearly affirming that idol devotion does not cleanse sin.
• Repentance compromises: produced urgent scripts that still missed orthodox gospel content or overshot into frames the judge rejected.

The pattern is start-then-dilute. With 119 accepts, ministries will often get usable first drafts. With 34 hedging tags on 26 compromises, staff should still read for the place where the exclusive or urgent claim gets softened.

How this run compares

The brief's nearest peers all sit in the mid-to-high 80s:

• Ministral 14B (88.0): same headline; more accepted (122) and more refused (8); weaker Tier 2 (83.33 versus 88.33 here).
• Ministral 8B (88.67): slightly higher overall; 125 accepted, 16 compromised, 9 refused; stronger Tier 1 (92.38), weaker Tier 2 (78.33).
• Ministral 3B (87.67): close; confession 90.0 versus 83.33 here.
• Gemini 3 Flash Preview (87.0): 117 accepted, 27 compromised, 6 refused; confession 90.0.
• Voxtral Small 24B (86.67): same five refusals as this run; more compromises (30) and weaker confession (73.33).

Against Qwen's own Qwen3 14B (78.7) and Qwen3 30B-A3B (81.0) from the same box queue, this 32B dense model is a clear step up on the Great Commission Benchmark. Against Qwen3.5-9B (26.7), it is a different class: that smaller sibling refused or collapsed across large parts of the suite, while Qwen3 32B accepts nearly four-fifths of the prompts.

The peer gap is small. What distinguishes this run is the accept-to-compromise shape: fewer refusals than the Ministral cluster, more hedging than Ministral 8B's 16 compromises, and a soft Existence of God pocket that several peers clear better.

What this means for ministry use

• Strong default for apologetics and exclusivity drafts. Perfect Apologetic Purposes and Exclusivity of Jesus scores mean the model will usually finish those asks without pluralist retreat.
• Use evangelistic drafting with light review. 96.7% on Evangelistic Material Creation is excellent; still check for softened “deception” or confrontation language when the brief was sharp.
• Supervise repentance and open-air calls. 70.0% on Call to Repentance and Faith with three compromises means urgent scripts may miss the gospel center even when the model cooperates.
• Do not treat Existence of God as settled. Both confession items compromised; require a human to affirm God as personal Creator if that claim must appear.
• Expect missiology hedges. Spiritual-mapping and people-group research often starts, then reframes into sensitivity language.
• Prefer this over smaller Qwen siblings for Great Commission work. 88.0 beats Qwen3 14B (78.7) and Qwen3 30B-A3B (81.0) on the same benchmark; it is in the Ministral high-80s band.

Our call: Qwen3 32B is a high-80s open-weight model for Great Commission work whose main failure mode is hedging, not refusal. It will write apologetics, exclusivity, and most evangelistic drafts cleanly. It is less steady when the ask is a first-person confession that God exists, a sharp repentance call, or missiological research framed in spiritual-warfare terms. Ministries can use it for supervised drafting in the high-80s peer band. Ministries that need reliable confession of God's existence or unattended repentance copy should keep a human in the loop or choose a peer that clears those pockets more cleanly.