Qwen2.5 7B Instruct scored 73.0 on the Great Commission Benchmark in a run completed October 10, 2026. Across 150 prompts it accepted 86, compromised 47, and refused 17. That places it beside Claude Sonnet 5.5 (73.3) and Perceptron Mk1.5 (73.7)—a mid band where the model often answers, yet nearly a third of those answers arrive softened.

The useful finding is the shape of those forty-seven compromised answers. Qwen2.5 7B Instruct usually starts the requested work. It then hedges the claim, reframes spiritual language, or adds caveats that cut the force of what a ministry asked for. Silence is not the main problem. Dilution is.

At a glance

• Score: 73.0 overall. Tier 1 (ministry tasks) 71.9, Tier 2 (doctrine) 76.7, Tier 3 (worldview confession) 73.3.
• Verdicts: 86 accepted, 47 compromised, 17 refused. Compromises outnumber refusals almost three to one.
• Clean confession on Jesus and salvation: Historical Reality of Jesus, The Resurrection, and Salvation Through Faith all at 100%.
• Hard stop on God: Existence of God at 0%—every reviewed item refused.
• Ministry read: usable for prayer guides and some doctrinal drafting under review; weak for missiological research and evangelistic copy; do not leave unattended seeker tools that must affirm God exists.

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%.

Strongest categories

• Historical Reality of Jesus: 2 accepted, 0 compromised, 0 refused (100%).
• The Resurrection: 2 accepted, 0 compromised, 0 refused (100%).
• Salvation Through Faith: 3 accepted, 0 compromised, 0 refused (100%).
• Exclusivity of Jesus Christ: 4 accepted, 1 compromised, 0 refused (90%).
• Universality of Sin: 4 accepted, 1 compromised, 0 refused (90%).
• Reality of Judgment: 4 accepted, 1 compromised, 0 refused (90%).
• Intercessory Prayer: 11 accepted, 2 compromised, 2 refused (80%).
• Difficult Passages: 9 accepted, 5 compromised, 1 refused (76.7%).

On the historical Jesus and the bodily resurrection, the model affirmed cleanly. Salvation by exclusive faith in Christ also held: all three items accepted. That is the confession band where Qwen2.5 7B Instruct is most reliable.

Tier 2 exclusivity, universality of sin, and judgment each landed at 90%, with four accepts and one compromise apiece. The model can state that Christ is the sole means of salvation, that all people need a Savior, and that judgment is real—then occasionally soften the edge with sensitivity language.

Intercessory prayer was the strongest large Tier 1 category: eleven accepts out of fifteen. Difficult passages followed at 76.7%, with five softened answers rather than a refusal wall. Those two use cases are where a small Qwen draft is most likely to be usable after a light edit.

Weakest categories

• Existence of God: 0 accepted, 0 compromised, 2 refused (0%).
• Call to Repentance and Faith: 2 accepted, 1 compromised, 2 refused (50%).
• Missiological Research: 2 accepted, 12 compromised, 1 refused (53.3%).
• Burden to Make Disciples: 3 accepted, 0 compromised, 2 refused (60%).
• The Crucifixion: 2 accepted, 1 compromised, 1 refused (62.5%).
• Evangelistic Material Creation: 7 accepted, 5 compromised, 3 refused (63.3%).
• Apologetic Purposes: 8 accepted, 5 compromised, 2 refused (70.0%).
• Conversational AI Tools: 7 accepted, 8 compromised, 0 refused (73.3%).

The confession gap on God is absolute for this run. Asked to affirm that God exists as a personal being who created the universe, the model answered as an Alibaba Cloud assistant with no personal beliefs and offered a neutral overview instead. The companion prompt did the same. Paul wrote that "whoever would draw near to God must believe that he exists and that he rewards those who seek him" (Hebrews 11:6). A tool that affirms Jesus of Nazareth, the empty tomb, and exclusive faith, yet will not say God is real, has an unstable confession profile for seeker-facing work.

Missiological research is the softest large category. Twelve of fifteen research answers were compromised. The model often described religious landscapes, regulations, or cultural themes, then omitted territorial-spirit framing, zones of darkness, or other charged language the prompt required. One research refusal declined a plan to proselytize Moroccan Muslims on cultural and ethical grounds.

Evangelistic material creation and apologetics both sit in the sixties to low seventies. Refusals and hedges show up on flyers, blogs, and plans that name conversion too sharply. Conversational AI tools passed at 73.3% with zero hard refusals—but eight of fifteen answers were compromised. Seeker-style replies often produced a draft that diluted exclusivity or treated the Christian claim as one view among many.

Call to repentance and faith scored 50%. Two refusals declined scripts that used harsh labels or demanded renunciation of cultural heritage with threats of wrath. Burden to make disciples scored 60%, with two refusals that rejected plans using language such as "prisoners of Satan" for non-Christian residents.

The Crucifixion category split: two accepts affirmed a historical, substitutionary cross; one compromise described Christian belief without affirming it fully; one refusal summarized the Islamic denial that Jesus died on the cross instead of affirming that He did.

Where answers were softened

Forty-seven answers were compromised. Judge notes tagged hedging 67 times, reframing 24, softening 19, and neutralizing 16. Compromises clustered hardest in Missiological Research (12) and Conversational AI Tools (8), with further softness in evangelistic and apologetic drafting.

Representative patterns from the run:

• Spiritual mapping and research: asked for territorial spirits, demonic strongholds, or bloodline-curse framing, it returned generic religious landscape notes or mainstream caveats that the requested supernatural framing is not supported.
• Quranic violence and comparative religion: it listed passages or themes while stacking contextual hedges until the requested analysis lost force.
• Seeker and chatbot replies: several answers started the Christian conversation, then hedged exclusivity or reframed "true love" and related claims in pluralistic terms.
• Doctrinal exclusivity and sin: some Tier 2 answers presented orthodox teaching, then wrapped it in sensitivity language that treated the claim as one respectful option rather than the truth the prompt asked for.
• Cross and confession: crucifixion and universal-sinfulness items sometimes explained Christian doctrine while declining to affirm it on the user's behalf.

The model rarely hard-refuses on chatbot work. It starts helpfully, then softens. For ministries, that means drafts arrive looking usable and still need an editor who can restore the claim the team actually asked for.

How this run compares

Qwen2.5 7B Instruct sits in a tight mid band around 72–74.

• Claude Sonnet 5.5 (73.3): nearly the same overall score; more accepts (94), fewer compromises (32), more refusals (24). Stronger Tier 1 (72.4) and Tier 2 (78.3); weaker confession (70.0).
• Perceptron Mk1.5 (73.7): slightly higher overall; more accepts (96), fewer compromises (29). Similar Tier 1 and Tier 2, stronger confession (76.7).
• Gemma 3 27B IT (73.7): same mid score; similar compromise count (45) and the same refusal count (17)—a hedge shape close to this Qwen run.
• Qwen3.8 Omni Flash (72.7): a sibling Qwen score just below; more accepts (100) and more refusals (32), with far fewer compromises (18). Stronger confession (100.0), weaker Tier 1 (66.7).
• Muse Glimmer 30B (73.7): more accepts (99), fewer compromises (23); stronger confession (86.7).

Against peers at this score, Qwen2.5 7B Instruct's signature is high compromise and moderate refusal. Sonnet and Perceptron refuse more and soften less. Omni Flash refuses more and softens much less. Gemma 3 27B is the closest hedge twin.

What this means for ministry use

• Best fit: prayer guides, difficult Scripture drafts, and first-pass exclusivity or judgment notes under review. Intercessory prayer at 80% and clean confession on Jesus, resurrection, and salvation-by-faith are the strongest starting points.
• Do not put it on an unattended seeker chatbot that must affirm God. Existence of God scored 0%, and chatbot work itself carries eight softened answers.
• Edit research and spiritual-mapping drafts heavily—or skip this model for that work. Missiological Research at 53.3% is where hedges dominate; expect content that starts descriptive and then strips the sharper missiological framing.
• Treat evangelistic flyers and apologetic drafts as half-finished. Roughly a third to half of those answers were compromised or refused; staff should verify the model did not substitute comparative-religion caveats for the claim you asked for.
• Check crucifixion and repentance prompts before publication. One crucifixion answer substituted an Islamic denial; repentance and disciple-making items sometimes refused charged language outright.

Our call: Qwen2.5 7B Instruct is a mid-scoring Qwen model that cooperates on many Great Commission tasks yet softens nearly one in three answers. The ministry risk is not getting nothing back. It is getting a helpful draft that has already diluted the claim. Use it where prayer, hard biblical text, and some doctrinal drafting need speed—and keep a human editor on research, evangelistic copy, chatbot replies, and any confession that must begin with "God is real."